Доменная адаптация глубоких сверточных нейросетей для обработки медицинских изображений / Domain Adaptation of Deep Convolutional Neural Networks in Medical Imaging тема диссертации и автореферата по ВАК РФ 00.00.00, кандидат наук Широких Борис Николаевич

• Широких Борис Николаевич
• кандидат наук
• 2026

Оглавление диссертации кандидат наук Широких Борис Николаевич

Contents

Page

Introduction

Chapter 1 Domain Shift Anatomy

1.1 Background

1.2 Methods

1.2.1 Predefined fine-tuning

1.2.2 Adaptive fine-tuning

1.3 Experiments

1.3.1 Data

1.3.2 Experimental setup

1.3.3 Evaluation metrics

1.3.4 Architecture and Training

1.4 Results

1.4.1 Baseline and Oracle

1.4.2 Predefined fine-tuning

1.4.3 Adaptive fine-tuning

1.5 Summary

Chapter 2 Mitigating Domain Shift from CT Reconstruction Kernels

2.1 Background

2.2 Methods

2.2.1 Filtered Back-Projection Augmentation

2.2.2 Cross-Domain Features Consistency

2.2.3 Cross-Domain Predictions Consistency

2.3 Data

2.3.1 Segmentation Data

2.3.2 Paired Images Data

2.4 Experiments

2.4.1 Segmentation model

2.4.2 FBPAug experiments and results

Page

2.4.3 F-consistency experiments and results

2.5 Summary

Chapter 3 Systematically Evaluating DA Methods in 3D Medical Image

Segmentation

3.1 Background

3.2 M3DA Benchmark

3.2.1 UDA Problems Motivation

3.2.2 Datasets Selection

3.3 Methods for M3DA Benchmark

3.3.1 Segmentation Baseline and Oracle

3.3.2 UDA Methods

3.4 Experiments

3.4.1 Backbone Selection

3.4.2 Domain Adaptation Methods on M3DA

3.4.3 Impact of Additional Augmentations

3.5 Discussion and Limitations

3.5.1 Alternative DA Frameworks

3.5.2 Limitations and Future Directions

3.6 Burdenko's Glioblastoma Progression Dataset

3.7 Summary

Chapter 4 Benchmark for OOD in 3D Medical Image Segmentation

4.1 Background

4.2 Data

4.2.1 Lung Nodules Segmentation

4.2.2 Vestibular Schwannoma Segmentation

4.2.3 Problem Setting

4.3 Methods

4.3.1 Methods Selection

4.3.2 Methods Implementation

4.3.3 Intensity Histogram Features

4.4 Experiments

4.4.1 Experimental Setup

Page

4.4.2 Results

4.4.3 In-Depth Benchmark Analysis

4.4.4 Ablation Study on Synthetic Data

4.5 Summary

Conclusion

List of symbols and abbreviations

List of Figures

List of Tables

Appendix A Certificate of Registration

Введение диссертации (часть автореферата) на тему «Доменная адаптация глубоких сверточных нейросетей для обработки медицинских изображений / Domain Adaptation of Deep Convolutional Neural Networks in Medical Imaging»

Introduction

The relevance of the research area. The degree of its development. The

rapid advancements in Deep Learning (DL) have established a strong foundation for automating medical image segmentation tasks [1]. Segmentation algorithms, particularly those based on Convolutional Neural Networks (CNNs), have achieved near human-level performance across a variety of clinical applications, including brain tumor, lung cancer, organ-at-risk, and liver tumor segmentation. These algorithms continue to evolve, offering increasingly accurate results.

A major focus of our research was to evaluate and integrate DL methods into real-world clinical workflows. One of the earliest examples is the deployment of an automatic segmentation system for stereotactic radiosurgery of brain metastases at the Moscow Gamma-Knife Center [157]. Radiosurgery demands submillimeter precision in target delineation, where even minor variability between specialists can affect the radiation dose distribution and, consequently, treatment outcomes. The developed CNN-based segmentation model significantly improved the clinical workflow by reducing inter-rater detection disagreements by 48%, enhancing contouring consistency (surface Dice score increased from 0.845 to 0.871), and halving the delineation time. The model was validated through a prospective clinical study involving four experts and twenty patients, showing its potential as a reliable assistant in radiosurgical planning. This work demonstrated that DL-based tools can yield measurable improvements in diagnostic precision and efficiency, highlighting their immediate clinical applicability and setting a methodological precedent for future clinical evaluations of DL systems.

Another practically oriented line of research addressed the COVID-19 pandemic, during which the radiology departments of Moscow healthcare institutions faced extreme diagnostic loads. In collaboration with the Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies, we developed a DL-based triage system for chest CT scans [158]. The proposed multitask CNN model jointly performed COVID-19 identification and severity quantification, effectively prioritizing patients for urgent care. Trained on approximately 1500 (mostly unannotated) CT studies and evaluated on an independent test set of 123 studies, the model achieved ROC AUC scores up to 0.97 for COVID-19 identification versus healthy controls and a 0.97 Spearman correlation for severity estimation. This system was successfully integrated

into the Moscow healthcare workflow [166], accelerating patient sorting and resource allocation during the pandemic.

These two implementations - one in precision radiosurgery, the other in large-scale public healthcare - demonstrate the high clinical relevance and transformative potential of DL methods in medical imaging.

Despite these successes, the widespread adoption of DL in medical imaging remains hindered by the significant drop in neural network performance when applied to data that differs from the training set. This challenge, known as domain shift, is particularly prevalent in medical imaging due to variations in scanner acquisition parameters, the introduction of new imaging modalities, and differences in patient populations. Even models that perform reliably in controlled or retrospective evaluations may degrade sharply once the imaging equipment or acquisition conditions change.

A practical instance of this problem was observed during the deployment of the Gamma-Knife segmentation system. In 2020, the MRI scanner at the clinic was replaced, resulting in visually distinct images with altered intensity distributions. The previously validated model, which had operated reliably for several years, began to produce a large number of false-positive contours, increasing the manual workload for radiologists instead of reducing it. Since the Gamma-Knife center treats approximately 400 patients per year and rare pathology cases were available only in the historical dataset (2005-2017), assembling a new annotated dataset of sufficient quality would require several years, which is impractical for clinical use. This situation motivated the study of supervised domain adaptation in the extreme low-data regime, aimed at adapting a deployed model when only a limited number of annotated target scans are available.

This clinical demand shaped the subsequent research direction, which we follow in our works [159; 160]. In this part of the study, we formalize the problem of supervised domain adaptation under conditions of severe annotation scarcity and analyze the internal sensitivity of CNNs to scanner-induced domain shifts. The analysis reveals that existing methods and architectures are not optimized for such low-data regimes and provide little understanding of which components of the model are most affected by distributional changes. Furthermore, the field lacks interpretable adaptation mechanisms and systematic approaches for identifying effective fine-tuning strategies. These limitations motivated the development of data-efficient and explainable supervised domain adaptation techniques.

A complementary set of deployment issues emerged while scaling the COVID-19 segmentation model across the Moscow healthcare network. When processing thousands of CT studies per day, the model encountered two major modes of failure: (i) systematic distributional changes caused by differing acquisition protocols - notably reconstruction kernels and radiation dose; and (ii) clearly out-of-distribution (OOD) inputs - such as head/abdomen scans routed into the chest stream, truncated volumes, or occasional synthetic/test images - that the model was never trained to handle. These failure modes expose the limits of purely supervised adaptation: a large unlabeled stream of data and rare but clinically critical OOD cases demand methods that can adapt from unlabeled data and reliably detect and reject anomalies before producing clinical predictions.

Building upon these observations, we formulate two complementary research directions. The first follows unsupervised domain adaptation (UDA) under CT reconstruction variability - acute but understudied case where changes in reconstruction kernels induce significant distributional shifts. While UDA is a well-established topic in general machine learning research [2], existing approaches lack the specificity and theoretical grounding required for medical imaging data [3-7]. Therefore, we develop theoretically grounded models in our works [161; 162], then investigate the specificity of UDA on volumetric medical images in the subsequent work [163]. The second direction addresses the problem of OOD detection in volumetric medical imaging. Although OOD detection methods have achieved strong results on natural images, they fail to scale effectively to 3D medical data, where both computational and statistical challenges are more severe [164].

These gaps motivate the development of specialized domain adptation (DA) and OOD detection methods specifically tailored for real-world clinical environments and 3D medical image segmentation in particular. Although the general principles of DA are well established in machine learning, existing approaches show limited robustness to clinically relevant domain shifts - such as differences in MRI intensity distributions and CT reconstruction kernels. Furthermore, the literature lacks standardized evaluation protocols that would enable systematic comparison of DA techniques in realistic medical scenarios.

This dissertation aims to fill these gaps by developing mathematically grounded and data-efficient DA methods, establishing comprehensive benchmarking frameworks, and publishing clinically relevant datasets that support reproducible evaluation. The proposed approaches address both supervised and unsupervised DA

settings, as well as OOD detection in 3D medical imaging, thereby contributing to the reliability and interpretability of DL models in clinical practice.

In summary, the dissertation consolidates theoretical development, empirical validation, and dataset curation to advance the field of domain adaptation in medical imaging and to bring DL models closer to robust, real-world deployment.

Goals and problems addressed. The primary goal of this thesis is to address the challenges posed by domain shifts in 3D medical image segmentation, which are a significant obstacle to the reliable and accurate performance of DL models in real-world clinical settings. By investigating various sources of domain shifts, particularly in MRI and CT images, the thesis aims to develop robust methods that improve the generalization of segmentation models across different imaging domains. This work not only focuses on enhancing the performance of models through developing fine-tuning and DA methods but also introduces novel benchmarks and a dataset to ensure that these models can effectively handle OOD data, which is critical for the safe deployment of deep learning (DL) in medical applications. To achieve the goal of the dissertation, the following problems are addressed:

1. Domain Shift Analysis in MRI and CT: Estimate the sensitivity of segmentation models to different sources of domain shift in MRI and CT images. Identify CNN layers most affected by domain shift and propose a strategies for automatic layer selection and interpretable domain adaptation.

2. Development of DA Algorithms: Design a theoretically grounded augmentation method to simulate CT kernel-induced shift. Develop a data-driven unsupervised adaptation method based on enforcing feature map similarity between paired CT images. Implement and evaluate supervised and unsupervised adaptation pipelines in low-data and zero-shot settings.

3. Development of the Benchmark for DA: Construct a large-scale benchmark for systematic evaluation of unsupervised DA methods in 3D medical image segmentation. Integrate multiple publicly available datasets to cover a diverse range of clinically relevant domain shift scenarios across imaging modalities, acquisition protocols, and contrast conditions. Identify key limitations of existing adaptation methods and provide unified evaluation protocols that enable fair comparison and reproducibility. Collect and publish a clinically realistic dataset for evaluating model robustness and adaptation under heterogeneous acquisition variability.

4. Benchmarking OOD Detection in 3D Medical Imaging: Develop a benchmark for OOD detection in 3D medical image segmentation, incorporating multiple clinically relevant test scenarios. Evaluate existing OOD detection approaches and identify their limitations when scaled to volumetric medical data. Propose a computationally efficient and interpretable OOD detection method.

Scientific novelty. The scientific novelty is built up from the following results:

1. A novel gradient-based method, SpotTUnet, was proposed and implemented to automatically identify and fine-tune the neural network layers most susceptible to domain shifts in supervised domain adaptation settings, improving segmentation performance under limited data availability.

2. A knowledge-driven augmentation technique, Filtered Back-Projection Augmentation (FBPAug), was developed for the first time to simulate the effect of CT reconstruction kernel variability, significantly improving segmentation consistency.

3. A data-driven unsupervised domain adaptation method, F-Consistency, was developed for the first time to address CT reconstruction kernel variability by enforcing feature map similarity between paired images. The method achieves state-of-the-art performance under kernel-induced domain shifts and establishes a foundation for future self-supervised learning approaches using FBPAug-generated pairs and the F-Consistency criterion.

4. A large-scale benchmark, M3DA, was introduced for evaluating unsupervised domain adaptation methods in 3D medical image segmentation, revealing critical limitations of existing approaches - none of which close more than 61% of the performance gap between source and target domains.

5. A new publicly available dataset was collected and published for primary glioblastoma segmentation in radiotherapy planning. It allows for testing domain adaptation methods in the close to clinical scenarios.

6. A benchmark for out-of-distribution detection in 3D medical imaging was constructed for the first time, containing multiple challenge scenarios and exposing critical limitations of existing OOD detection methods.

Theoretical and practical significance. From a theoretical perspective, the dissertation contributes novel formulations and learning principles for domain adaptation in 3D medical image segmentation. The proposed SpotTUnet method introduces a gradient-based criterion for identifying domain-shift-sensitive layers in CNNs, offering new insights into the structural behavior of deep models under

distributional shift. The FBPAug method is grounded in the mathematics of computed tomography, providing a parametrically controlled augmentation technique to simulate domain shifts caused by reconstruction kernels. F-Consistency formulates an unsupervised learning objective based on the similarity of internal representations, establishing a principled approach for leveraging paired but unlabeled data. These developments enrich the foundations of machine learning, domain adaptation in particular, and support further study into robustness and model generalization.

On the practical side, the proposed methods have been shown to significantly improve segmentation robustness under clinically relevant domain shifts, including those caused by acquisition parameters, reconstruction algorithms, and institutional variability. The developed M3DA benchmark provides a comprehensive framework for empirical evaluation of DA algorithms, while the published BGP dataset enables realistic testing in radiotherapy planning scenarios. Additionally, the OOD detection benchmark and the IHF method contribute to building reliable diagnostic pipelines capable of identifying OOD inputs in practice.

Altogether, the results of this dissertation contribute to advancing the safe deployment of DL in clinical practice and help bridge the gap between algorithmic development and medical application.

The majority of methods and benchmarks have publicly available source code:

1. First-layers fine-tuning (author-B. Shirokikh): https://github.com/kechua/DART20

2. Domain Shift Anatomy (co-authored -1. Zakazov, B. Shirokikh): https://github.com/neuro-ml/domain_shift_anatomy

3. FBPAug (co-authored; software development - T. Saparov): https://github.com/STNLd2/FBPAug

4. M3DA benchmark (author-B. Shirokikh): https://github.com/BorisShirokikh/M3DA

5. BGP Dataset (co-authored; data processing pipeline — B. Shirokikh): cancerimagingarchive.net/collection/burdenko-gbm-progression/

6. Benchmark for Out-of-distribution detection on 3D medical images (co-authored; benchmark design and evaluation — B. Shirokikh): https://github.com/francisso/OOD-benchmark

7. Intensity Histogram Features algorithm (author - B. Shirokikh): github.com/BorisShirokikh/MOOD_submission_Sample-level_AIRI

The listed software products were developed by the author personally or with his direct participation. The author's contribution includes the formulation of methods and algorithms, design of experimental pipelines, and, where applicable, software implementation. In joint works, software development was carried out collaboratively, with co-authorship explicitly reflected in the corresponding publications.

In addition to being published and open-sourced, FBPAug has been integrated into multiple products at IRA Labs Ltd, including a COVID-19 segmentation algorithm registered as Intellectual Property1, further demonstrating the practical significance of this work. The corresponding certificate of state registration of computer programs could be found in Appendix A.

Research methodology. The research employs machine learning and deep learning techniques to improve the generalization and robustness of CNNs in 3D medical image segmentation, focusing on both supervised and unsupervised learning paradigms. Methodologically, the work integrates tools from linear algebra, analysis, probability theory, mathematical statistics, numerical optimization, software development, and mathematical models of computer tomography to develop and analyze domain adaptation algorithms.

Experimental evaluation relies on large-scale empirical studies using publicly available and newly introduced benchmarks and datasets. All methods are implemented in the Python programming language using established DL frameworks. Software components were designed for reproducibility and extensibility, supporting broader use in medical image analysis. Clinical relevance and task formulation were refined through collaboration with medical experts in radiology and radiation oncology.

Propositions submitted for defense.

1. A novel supervised domain adaptation method, SpotTUnet, was developed to automatically identify and fine-tune layers most affected by domain shift. The method improved segmentation performance in MRI across diverse clinical domains and provided interpretable insights into layer-wise shift susceptibility. Used as a visualization tool, SpotTUnet contributed to the development of the other proposed methods, such as F-Consistency and IHF.

2. A knowledge-driven augmentation technique, FBPAug, was proposed to address domain shifts caused by different CT reconstruction kernels. The method significantly increased prediction consistency across paired images, improving Dice score from 0.46 to 0.76 in COVID-19 lung segmentation.

1https://www.fips.ru/registers-doc-view/fips_servlet?DB=EVM&DocNumber=2021612647

Furthermore, FBPAug was used to develop most of the commercially used CT segmentation algorithms, significantly improving their robustness.

3. A data-driven unsupervised adaptation method, F-Consistency, was introduced to leverage paired CT images reconstructed with different kernels. By enforcing feature map similarity, the method achieved state-of-the-art results in kernel-induced domain shift settings, increasing Dice score to 0.80.

4. The M3DA benchmark was constructed for the large-scale evaluation of unsupervised DA methods in 3D medical image segmentation. It comprises four publicly available datasets and eight clinically relevant domain shift setups, including shifts across modalities, acquisition protocols, and contrast settings.

5. A new dataset, Burdenko's Glioblastoma Progression (BGP), was collected and published to evaluate segmentation models in a realistic DA scenario. It includes multi-sequence MRI scans from 180 patients acquired across multiple clinical sites and scanner vendors, enabling testing under heterogeneous acquisition variability.

6. A benchmark for OOD detection in 3D medical segmentation was proposed, containing several clinically relevant challenge cases. It revealed fundamental limitations of existing OOD methods and established practical evaluation criteria for real-world robustness assessment.

7. An effective OOD detection method, Intensity Histogram Features (IHF), was developed. With the lowest computational cost, it achieved top-2 ranking in the Medical Out-of-Distribution Challenge (MOOD) 2022 and 2023, making it a strong baseline and analysis tool for domain shifts in 3D medical images.

Validity of the obtained results. Approbation. The statements and conclusions

formulated in the dissertation have received qualified approbation at the following international and Russian scientific conferences:

1. Domain Adaptation and Representation Transfer MICCAI 2020 Workshop. October 2020. Lima, Peru.

2. Cross-Modality Domain Adaptation for Medical Image Segmentation (challenge-associated satellite event at MICCAI). September 2021. Strasbourg, France.

3. Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI). September 2021. Strasbourg, France.

4. Information Technologies and Systems (ITaS) conference. November 2021. Moscow, Russian Federation.

5. Medical Out-of-Distribution Analysis Challenge (challenge-associated satellite event at MICCAI). September 2022. Singapore, Republic of Singapore.

6. Information Technologies and Systems (ITaS) conference. September 2023. Istra, Russian Federation.

The reliability of the results is ensured by comprehensive experimental evaluation across multiple datasets, diverse experimental settings, and multiple methods of statistical analysis, with all validation procedures in accordance with the leading work in the area. The reproducibility of findings was validated through systematic benchmarking of multiple deep learning algorithms and repeated experiments to ensure statistical significance of the results. All the proposed methods are compared with existing methods for domain adaptation or OOD detection using measurable metrics.

The credibility is also confirmed by two publications of research results in Q1 peer-reviewed scientific journals, two publications in Q2 peer-reviewed journals, and three peer-reviewed conference proceedings, including one in proceedings of the Rank A conference. All of these journals and proceedings are indexed in Scopus and Web of Science.

Personal contribution of the author. The content of the dissertation and the main statements submitted for defense reflect the author's personal contribution to the published works, obtained personally by the author or with his direct participation. The author's contribution includes the formulation of research objectives, literature analysis, design and execution of extensive computational experiments, and interpretation of the obtained results. The dissertation aims and objectives were formulated jointly by the author, the author's scientific supervisor, Prof. Ivan V. Oseledets, and Dr. Mikhail G. Belyaev (IRA Labs Ltd).

Preparation of the seven publications (four in peer-reviewed journals and three in conference proceedings) was carried out together with co-authors, with the author's contribution being decisive. In particular, the author proposed the core methodologies for domain adaptation and out-of-distribution detection, developed experimental pipelines, and analyzed experimental results. Additionally, the author contributed to the development of a software product registered by the state and to the creation of a publicly available medical imaging dataset.

Dissertation structure. The dissertation consists of an abstract, introduction, 4 chapters, conclusions, lists of abbreviations, figures, tables, and 1 appendix. The full volume of the dissertation is 132 pages with 24 figures and 15 tables. The list of references contains 166 sources.

Contents of the Dissertation. The Introduction substantiates the relevance of the work, formulates the research goals and problems, outlines the scientific novelty, and describes the methodology and practical significance of the results. Chapter 1 analyzes the internal sensitivity of convolutional neural networks to domain shifts in MRI segmentation. A gradient-based layer selection method (SpotTUnet) is proposed, allowing autonomous and interpretable fine-tuning under data-limited conditions. Chapter 2 addresses domain shifts arising from reconstruction kernel variability. Two novel approaches are developed: a theoretically grounded augmentation method (FBPAug) and an unsupervised adaptation framework (F-Consistency), both demonstrating state-of-the-art performance and practical applicability in clinical imaging. Chapter 3 introduces the large-scale M3DA benchmark for evaluating unsupervised domain adaptation in 3D medical image segmentation and presents the BGP dataset, offering a realistic "in-the-wild" validation scenario. Chapter 4 develops an evaluation framework for OOD detection in medical images and proposes a computationally efficient method, IHF, which achieves top-ranked results in international challenges. The Conclusion summarizes the main scientific results, their theoretical and practical significance, limitations, and outlines prospects for future research.

Заключение диссертации по теме «Другие cпециальности», Широких Борис Николаевич

Conclusion

This dissertation is devoted to the development of mathematically grounded and practically robust deep learning methods for domain adaptation and out-of-distribution detection in 3D medical image segmentation. The work systematically addresses the challenge of domain shift, which remains a critical obstacle to the reliable deployment of machine learning systems in clinical environments.

1. A gradient-based supervised domain adaptation method, SpotTUnet, was proposed to address the problem of layer-wise adaptation in convolutional neural networks for medical image segmentation. The method optimizes a learnable layer selection policy, allowing to fine-tune the most shift-sensitive layers. The effectiveness of this approach was confirmed through extensive experimental validation, including statistically significant improvements over common fine-tuning baselines in few-shot adaptation scenarios. From a practical standpoint, SpotTUnet provides interpretable visualizations of domain shift sensitivity across network layers, which we further used to guide the design of F-Consistency and to enhance performance of existing DA algorithms, such as DANN [2], on the M3DA benchmark. Moreover, conclusions drawn from SpotTUnet contributed to the development of the IHF method for out-of-distribution detection.

2. Two complementary DA methods were proposed to address performance degradation in CT segmentation caused by variations in reconstruction kernels. First, a knowledge-driven augmentation technique, Filtered Back-Projection Augmentation (FBPAug), was developed to simulate such kernel-induced domain shifts by modeling the mathematics of CT image formation in sinogram space. Second, a data-driven unsupervised DA method, F-Consistency, was introduced to align internal network representations of paired CT images reconstructed with different kernels. Both methods demonstrated statistically significant improvements over state-of-the-art approaches: FBPAug achieved high consistency in the zero-shot adaptation setting, while F-Consistency further increased segmentation quality in the standard unsupervised DA setting. FBPAug has been integrated into the production pipeline of a medical imaging startup, where it significantly

improved the robustness and accuracy of CT-based segmentation models used in clinical practice.

3. A large-scale benchmark, M3DA, was developed to evaluate unsupervised DA methods in 3D medical image segmentation under realistic and diverse shift scenarios. The benchmark reveals that existing methods close only 62% of the performance gap on average. To support clinical relevance, a new multi-modal dataset, BGP, was published for glioblastoma segmentation with high intra-institution variability. These contributions enable systematic comparison of methods and guide the design of robust and generalizable DA algorithms.

4. A benchmark for OOD detection in 3D medical segmentation was constructed, including clinically relevant scenarios, and was used to identify fundamental limitations of current methods. A lightweight method, IHF, was proposed as an interpretable and computationally efficient baseline, achieving top-2 in the MOOD 2022 and 2023 challenges. The benchmark and IHF serve as a foundation for further theoretical study of distributional uncertainty in medical imaging data.

The results and insights of this dissertation have already served as a foundation for several subsequent scientific works. Firstly, we used the proposed OOD benchmark to develop a novel OOD detection metric and theoretically justified framework [152]. We also used components of this benchmark to analyze predictive uncertainty at the level of distinct predicted components [153]. Finally, several colleagues' papers [154-156] have already complemented our research by integrating FBPAug into a self-supervised learning (SSL) framework, realizing a direction for augmentation that was originally suggested in this dissertation. These works demonstrate the relevance of the presented contributions and point toward continued progress in the robust deployment of deep learning in medical imaging.

While the dissertation advances the state of domain adaptation and out-of-distribution detection in 3D medical image segmentation, several limitations define promising directions for future research.

First, SpotTUnet currently operates in a supervised DA setup and relies on annotated target samples to optimize its layer-wise adaptation policy. Extending this method to an unsupervised setting - by introducing auxiliary domain shift criteria, e.g., domain adversarial loss function, - could substantially increase its applicability in other experimental settings. Furthermore, as transformer-based architectures [90]

increasingly replace convolutional backbones in segmentation models, developing analogous interpretable adaptation strategies for these architectures becomes an important task.

Second, although FBPAug and F-Consistency address the CT reconstruction kernel domain shift from complementary theoretical perspectives, they remain independent methods. A natural next step is to integrate FBPAug-generated synthetic pairs within the F-Consistency framework, thereby combining knowledge-driven and data-driven adaptation strategies into a unified self-supervised learning (SSL) approach, that may result in the inherently robust CT segmentation model. The follow-up work has already included FBPAug as an essential augmentation method into their contrastive SSL pretraining frameworks [154; 155].

Third, future research should systematically evaluate emerging SSL and diffusion-based models within the M3DA and BGP benchmarks, as well as develop SSL paradigms explicitly aware of domain shifts in loss formulation. In particular, DA-aware contrastive objectives may provide a principled way to learn domain-invariant yet semantically consistent representations.

Fourth, future work in the field of OOD detection could extend the developed state-of-the-art IHF algorithm from image-level to a pixel-level one, resulting in a segmentation-level detection of anomalies. Recent progress in SSL-based anomaly localization [156] highlights the feasibility of this direction, suggesting that combining pixel-wise uncertainty estimation with feature-space distribution modeling could substantially improve model interpretability and safety in clinical use.

Overall, the presented research constitutes a systematic investigation into the reliability and robustness of deep learning models for 3D medical image segmentation under domain shift. By introducing theoretically grounded methods, publicly available benchmarks, and practically validated algorithms, this dissertation bridges the gap between academic methodology and clinical application. The results contribute to the broader effort of building trusted machine learning in medicine - systems that not only achieve high accuracy but also remain consistent, interpretable, and safe under real-world variability. These foundations pave the way toward the next generation of adaptive and self-aware medical artificial intelligence systems.

Список литературы диссертационного исследования кандидат наук Широких Борис Николаевич, 2026 год

Bibliography

1. J.-G. Lee [et al.] / Deep learning in medical imaging: general overview // Korean journal of radiology. — 2017. — Vol. 18, Is. 4. — P. 570-584.

2. Y. Ganin, V. Lempitsky / Unsupervised domain adaptation by backpropagation

// International conference on machine learning. — 2015. — P. 1180-1189.

3. I. Gulrajani, D. Lopez-Paz / In search of lost domain generalization [Electronic Resource]. — 2020. — URL: https://arxiv.org/abs/2007.01434 (visited on 10/20/2025).

4. G. Wilson, D. J. Cook / A survey of unsupervised deep domain adaptation //

ACM Transactions on Intelligent Systems and Technology (TIST). — 2020. — Vol. 11, Is. 5.—P. 1-46.

5. F. Zhuang [et al.] / A comprehensive survey on transfer learning // Proceedings of the IEEE. — 2020. — Vol. 109, Is. 1. — P. 43-76.

6. X. Peng [et al.] / Visda: A synthetic-to-real benchmark for visual domain adaptation // Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. — 2018. — P. 2021-2026.

7. H. Zhang [et al.] / An empirical framework for domain generalization in clinical settings // Proceedings of the conference on health, inference, and learning. — 2021.—P. 279-290.

8. D. Shen, G. Wu, H.-I. Suk / Deep learning in medical image analysis // Annual review of biomedical engineering. — 2017. — Vol. 19, Is. 1. — P. 221-248.

9. M. Wang, W. Deng / Deep Visual Domain Adaptation: A Survey // Neurocomputing. — 2018. — Vol. 312. — P. 135-153.

10. K. Kamnitsas [et al.] / Unsupervised domain adaptation in brain lesion segmentation with adversarial networks // International conference on information processing in medical imaging. — 2017. — P. 597-609.

11. B. Glocker [et al.] / Machine Learning with Multi-Site Imaging Data: An Empirical Study on the Impact of Scanner Effects [Electronic Resource]. — 2019. — URL: https://arxiv.org/abs/1910.04597 (visited on 10/20/2025).

12. M. Orbes-Arteaga [et al.] / Multi-domain adaptation in brain MRI through paired consistency and adversarial learning // Domain Adaptation and Representation Transfer and Medical Image Learning with Less Labels and Imperfect Data. — 2019.—P. 54-62.

13. J. Yosinski [et al.] / How transferable are features in deep neural networks? //

Advances in neural information processing systems. — 2014. — Vol. 27.

14. K. Kushibar [et al.] / Supervised domain adaptation for automatic sub-cortical brain structure segmentation with minimal user interaction // Scientific reports. — 2019. — Vol. 9, Is. 1. — P. 6742.

15. V. V. Valindria [et al.] / Domain adaptation for MRI organ segmentation using reverse classification accuracy [Electronic Resource]. — 2018. — URL: https: //arxiv.org/abs/1806.00363 (visited on 10/20/2025).

16. S. Valverde [et al.] / One-shot domain adaptation in multiple sclerosis lesion segmentation using convolutional neural networks // Neurolmage: Clinical. — 2019.— Vol. 21.—P. 101638.

17. M. Ghafoorian [et al.] / Transfer learning for domain adaptation in MRI: Application in brain lesion segmentation // International conference on medical image computing and computer-assisted intervention. — 2017. — P. 516-524.

18. K. Lee [et al.] / A simple unified framework for detecting out-of-distribution samples and adversarial attacks // Advances in neural information processing systems. — 2018. — Vol. 31.

19. E. Erdil, K. Chaitanya, E. Konukoglu / Unsupervised out-of-distribution detection using kernel density estimation [Electronic Resource]. — 2020. — URL: https://arxiv.org/abs/2006.10712 (visited on 10/20/2025).

20. R. Aljundi, T. Tuytelaars / Lightweight unsupervised domain adaptation by convolutional filter reconstruction // European Conference on Computer Vision. — 2016. — P. 508-515.

21. N. Karani [et al.] / Test-time adaptable neural networks for robust medical image segmentation // Medical Image Analysis. — 2021. — Vol. 68. — P. 101907.

22. Q. Dou [et al.] / Unsupervised Cross-Modality Domain Adaptation of Convnets for Biomedical Image Segmentations with Adversarial Loss // Proceedings of the 27th International Joint Conference on Artificial Intelligence. — 2018. — P. 691-697.

23. X. Zhao [et al.] / Robust white matter hyperintensity segmentation on unseen domain // 2021 IEEE 18th International Symposium on Biomedical Imaging (ISBI).— 2021.— P. 1047-1051.

24. Y. Guo [et al.] / Spottune: transfer learning through adaptive fine-tuning //

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. — 2019. — P. 4805-4814.

25. O. Ronneberger, P. Fischer, T. Brox / U-net: Convolutional networks for biomedical image segmentation // International Conference on Medical image computing and computer-assisted intervention. — 2015. — P. 234-241.

26. K. He [et al.] / Deep residual learning for image recognition // Proceedings of the IEEE conference on computer vision and pattern recognition. — 2016. — P. 770-778.

27. F. Milletari, N. Navab, S.-A. Ahmadi / V-net: Fully convolutional neural networks for volumetric medical image segmentation// 2016fourth international conference on 3D vision (3DV). — 2016. — P. 565-571.

28. A. Veit, M. J. Wilber, S. Belongie / Residual networks behave like ensembles of relatively shallow networks // Advances in neural information processing systems. — 2016. — Vol. 29. — P. 550-558.

29. E. Jang, S. Gu, B. Poole / Categorical Reparametrization with Gumble-Softmax [Electronic Resource]. — 2016. — URL: https://arxiv.org/abs/1611.01144 (visited on 10/29/2025).

30. R. Souza [et al.] / An open, multi-vendor, multi-field-strength brain MR dataset and analysis of publicly available skull stripping methods agreement // NeuroImage. — 2018. — Vol. 170. — P. 482-494.

31. S. Bakas [et al.] / Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the BRATS challenge [Electronic Resource]. — 2018. — URL: https://arxiv. org/abs/1811.02629 (visited on 10/20/2025).

32. S. Nikolov [et al.] / Clinically applicable segmentation of head and neck anatomy for radiotherapy: deep learning algorithm development and validation study // Journal of medical Internet research. — 2021. — Vol. 23, Is. 7. — P. e26151.

33. Ö. Çiçek [et al.] / 3D U-Net: learning dense volumetric segmentation from sparse annotation // International conference on medical image computing and computer-assisted intervention. — 2016. — P. 424-432.

34. I. Zacharov [et al.] / 'Zhores' - Petaflops supercomputer for data-driven modeling, machine learning and artificial intelligence installed in Skolkovo Institute of Science and Technology // Open Engineering. — 2019. — Vol. 9. — P. 512-520.

35. R. Schofield [et al.] / Image reconstruction: Part 1-understanding filtered back projection, noise and image acquisition // Journal of cardiovascular computed tomography. — 2020. — Vol. 14, Is. 3. — P. 219-225.

36. S. Schaller [et al.] / Spatial domain filtering for fast modification of the tradeoff between image sharpness and pixel noise in computed tomography // IEEE transactions on medical imaging. — 2003. — Vol. 22, Is. 7. — P. 846-853.

37. J. R. Zech [et al.] / Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study // PLoS medicine. — 2018. — Vol. 15, Is. 11. — P. e1002683.

38. M. Kloenne [et al.] / Domain-specific cues improve robustness of deep learning-based segmentation of CT volumes // Scientific reports. — 2020. — Vol. 10, Is. 1.—P. 1-9.

39. J. Choe [et al.] / Deep learning-based image conversion of CT reconstruction kernels improves radiomics reproducibility for pulmonary nodules or masses // Radiology. — 2019. — Vol. 292, Is. 2. — P. 365-373.

40. S. M. Lee [et al.] / CT image conversion among different reconstruction kernels without a sinogram by using a convolutional neural network // Korean journal of radiology. — 2019. — Vol. 20, Is. 2. — P. 295.

41. V. Sandfort [et al.] / Data augmentation using generative adversarial networks (CycleGAN) to improve generalizability in CT segmentation tasks // Scientific reports. — 2019. — Vol. 9, Is. 1. — P. 1-9.

42. A. D. Missert [et al.] / Simulation of CT images reconstructed with different kernels using a convolutional neural network and its implications for efficient CT workflow // Medical Imaging 2019: Physics of Medical Imaging. — 2019. — Vol. 10948.—P. 109482Y.

43. S. P. Morozov [et al.] / A simplified cluster model and a tool adapted for collaborative labeling of lung cancer CT scans // Computer Methods and Programs in Biomedicine. — 2021. — Vol. 206. — P. 106111.

44. S. Morozov [et al.] / MosMedData: data set of 1110 chest CT scans performed during the COVID-19 epidemic // Digital Diagnostics. — 2020. — Vol. 1, Is. 1. —P. 49-59.

45. E. B. Tsai [et al.] / The RSNA International COVID-19 Open Radiology Database (RICORD) //Radiology. —2021. — Vol. 299, Is. 1. — P. E204-E213.

46. S. G. Armato III [et al.] / The lung image database consortium (LIDC) and image database resource initiative (IDRI): a completed reference database of lung nodules on CT scans // Medical physics. — 2011. — Vol. 38, Is. 2. — P. 915-931.

47. K. J. Kiser [et al.] / Data from the Thoracic Volume and Pleural Effusion Segmentations in Diseased Lungs for Benchmarking Chest CT Processing Pipelines (PleThora) [Data set] [Electronic Resource]. — 2020. — URL: https: //doi.org/10.7937/tcia.2020.6c7y-gq39 (visited on 10/29/2025).

48. H. J. W. L. Aerts [et al.] / Data From NSCLC-Radiomics (version 4) [Data set] [Electronic Resource]. — 2014. — URL: https://doi.org/10.7937/K9/TCIA. 2015.PF0M9REI (visited on 10/30/2025).

49. D. P. Kingma / Adam: A method for stochastic optimization [Electronic Resource]. — 2014. — URL: https://arxiv.org/abs/1412.6980 (visited on 10/20/2025).

50. A. Tureckova [et al.] / Improving CT image tumor segmentation through deep supervision and attentional gates // Frontiers in Robotics and AI. — 2020. — Vol. 7.—P. 106.

51. J. Jiang, H. Veeraraghavan / Unified cross-modality feature disentangler for unsupervised multi-domain MRI abdomen organs segmentation // International conference on medical image computing and computer-assisted intervention. — 2020.—P. 347-358.

52. Q. Yu [et al.] / Source-Free Domain Adaptation for Medical Image Segmentation via Prototype-Anchored Feature Alignment and Contrastive Learning //

International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2023. — P. 3-12.

53. H. Zheng [et al.] / Hierarchical self-supervised learning for medical image segmentation based on multi-domain data aggregation // International conference on medical image computing and computer-assisted intervention. — 2021.—P. 622-632.

54. Q. Liu, Q. Dou, P.-A. Heng / Shape-aware meta-learning for generalizing prostate MRI segmentation to unseen domains // International conference on medical image computing and computer-assisted intervention. — 2020. — P. 475-485.

55. C. Chen [et al.] / Maxstyle: Adversarial style composition for robust medical image segmentation // International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2022. — P. 151-161.

56. R. Gu [et al.] / Domain composition and attention for unseen-domain generalizable medical image segmentation // International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2021. — P. 241-250.

57. J. Lennartz, T. Schultz / Segmentation Distortion: Quantifying Segmentation Uncertainty Under Domain Shift via the Effects of Anomalous Activations //

International Conference on Medical Image Computing and Computer-Assisted Intervention. —2023. —P. 316-325.

58. C. S. Perone [et al.] / Unsupervised domain adaptation for medical imaging segmentation with self-ensembling // NeuroImage. — 2019. — Vol. 194. — P. 1-11.

59. X. Han [et al.] / Deep symmetric adaptation network for cross-modality medical image segmentation // IEEE transactions on medical imaging. — 2021. — Vol. 41, Is. 1.—P. 121-132.

60. R. Dorent [et al.] / CrossMoDA 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation // Medical Image Analysis. — 2023. — Vol. 83. — P. 102628.

61. D. Al Chanti, D. Mateus / OLVA: O ptimal L atent V ector A lignment for Unsupervised Domain Adaptation in Medical Image Segmentation //

International Conference on Medical Image Computing and Computer-Assisted Intervention. —2021. —P. 261-271.

62. C. Weihsbach [et al.] / DG-TTA: Out-of-domain medical image segmentation through Domain Generalization and Test-Time Adaptation [Electronic Resource]. — 2023. — URL: https : // arxiv. org / abs / 2312 . 06275 (visited on 10/20/2025).

63. K. C. Wong, H. Wang, T. Syeda-Mahmood / HartleyMHA: Self-attention in Frequency Domain for Resolution-Robust and Parameter-Efficient 3D Image Segmentation // International Conference on Medical Image Computing and Computer-Assisted Intervention. —2023. —P. 364-373.

64. X. Liu [et al.] / Act: Semi-supervised domain-adaptive medical image segmentation with asymmetric co-training // International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2022. — P. 66-76.

65. C.-F. Cao [et al.] / CT scans and cancer risks: a systematic review and dose-response meta-analysis // BMC cancer. — 2022. — Vol. 22, Is. 1. — P. 1238.

66. D. J. Brenner, E. J. Hall / Computed tomography—an increasing source of radiation exposure // New Englandjournal of medicine. — 2007. — Vol. 357, Is. 22. — P. 2277-2284.

67. D. Nie [et al.] / Estimating CT image from MRI data using 3D fully convolutional networks // International Workshop on Deep Learning in Medical Image Analysis. — 2016. — P. 170-178.

68. V. R. Paczona [et al.] / Magnetic Resonance Imaging-Based Delineation of Organs at Risk in the Head and Neck Region // Advances in Radiation Oncology. — 2023. — Vol. 8, Is. 2. — P. 101042.

69. J. Vymazal [et al.] / Comparison of CT and MR imaging in ischemic stroke // Insights into imaging. — 2012. — Vol. 3. — P. 619-627.

70. T. Kubo [et al.] / Standard-dose vs. low-dose CT protocols in the evaluation of localized lung lesions: Capability for lesion characterization—iLEAD study //

European journal of radiology Open. — 2016. — Vol. 3. — P. 67-73.

71. D. Raghavan [et al.] / Initial results from mobile low-dose computerized tomographic lung cancer screening unit: improved outcomes for underserved populations // The oncologist. — 2020. — Vol. 25, Is. 5. — P. e777-e781.

72. M. Mohammadshahi [et al.] / Cost effectiveness of mobile versus fixed computed tomography and magnetic resonance imaging: a systematic review // Iranian Journal of Public Health. — 2019. — Vol. 48, Is. 8. — P. 1418.

73. M. Andreucci, R. Solomon, A. Tasanarong, [et al.] / Side effects of radiographic contrast media: pathogenesis, risk factors, and prevention // BioMed research international. — 2014. — Vol. 2014.

74. C. M. Costelloe, B. Amini, J. E. Madewell / Risks and benefits of gadolinium-based contrast-enhanced MRI // Seminars in Ultrasound, CT and MRI. — 2020. — Vol. 41, Is. 2. — P. 170-182.

75. W. Yan [et al.] / MRI manufacturer shift and adaptation: increasing the generalizability of deep learning segmentation for MR images acquired with different scanners // Radiology: Artificial Intelligence. — 2020. — Vol. 2, Is. 4. —P. e190195.

76. Y. Li, Y. Fan / Medical Image Segmentation with Domain Adaptation: A Survey [Electronic Resource]. — 2023. — URL: https://arxiv.org/abs/2311.01702 (visited on 10/20/2025).

77. H. J. Kuijf [et al.] / Standardized assessment of automatic segmentation of white matter hyperintensities and results of the WMH segmentation challenge // IEEE transactions on medical imaging. — 2019. — Vol. 38, Is. 11. —P. 2556-2568.

78. B. H. Menze [et al.] / The multimodal brain tumor image segmentation benchmark (BRATS) // IEEE transactions on medical imaging. — 2014. — Vol. 34, Is. 10. — P. 1993-2024.

79. O. Bernard [et al.] / Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved? // IEEE transactions on medical imaging. — 2018. — Vol. 37, Is. 11. —P. 2514-2525.

80. V. M. Campello [et al.] / Multi-centre, multi-vendor and multi-disease cardiac segmentation: the M&Ms challenge // IEEE Transactions on Medical Imaging. — 2021. — Vol. 40, Is. 12. — P. 3543-3554.

81. F. Prados [et al.] / Spinal cord grey matter segmentation challenge // Neuroimage. — 2017. — Vol. 152. — P. 312-329.

82. G. Zheng, D. Belavy, S. Li / IVDM3Seg Dataset [Electronic Resource]. — 2018. — URL: https://ivdm3seg.weebly.com/data.html (visited on 10/20/2025).

83. B. Landman [et al.] / Miccai multi-atlas labeling beyond the cranial vault-workshop and challenge // Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge. — 2015. — Vol. 5. —P. 12.

84. Y. Ji [et al.] / Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation // Advances in Neural Information Processing Systems. — 2022. — Vol. 35. — P. 36722-36732.

85. A. E. Kavur [et al.] / CHAOS challenge-combined (CT-MR) healthy abdominal organ segmentation // Medical Image Analysis. — 2021. — Vol. 69. — P. 101950.

86. X. Zhuang, J. Shen / Multi-scale patch and multi-modality atlases for whole heart segmentation of MRI //Medical image analysis. — 2016. — Vol. 31. — P. 77-87.

87. C. P. Purysko, R. Renapurkar, M. A. Bolen / When does chest CT require contrast enhancement? // Cleveland Clinic Journal of Medicine. — 2016. — Vol. 83, Is. 6. —P. 423-426.

88. F. Isensee [et al.] / nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation // Nature methods. — 2021. — Vol. 18, Is. 2. — P. 203-211.

89. F. Isensee [et al.] / nnu-net revisited: A call for rigorous validation in 3d medical image segmentation // International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2024. — P. 488-498.

90. A. Hatamizadeh [et al.] / Unetr: Transformers for 3d medical image segmentation

// Proceedings of the IEEE/CVF winter conference on applications of computer vision. — 2022. — P. 574-584.

91. A. Hatamizadeh [et al.] / Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images // International MICCAI Brainlesion Workshop. — 2021. — P. 272-284.

92. S. Roy [et al.] / Mednext: transformer-driven scaling of convnets for medical image segmentation // International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2023. — P. 405-415.

93. H. Guan, M. Liu / Domain adaptation for medical image analysis: a survey // IEEE Transactions on Biomedical Engineering. — 2021. — Vol. 69, Is. 3. — P. 1173-1185.

94. M. Toldo [et al.] / Unsupervised domain adaptation in semantic segmentation: a review // Technologies. — 2020. — Vol. 8, Is. 2. — P. 35.

95. M. Ghifary, W. B. Kleijn, M. Zhang / Domain adaptive neural networks for object recognition // Pacific Rim international conference on artificial intelligence. — 2014.—P. 898-904.

96. E. Tzeng [et al.] / Deep domain confusion: Maximizing for domain invariance [Electronic Resource]. — 2014. — URL: https://arxiv.org/abs/1412.3474 (visited on 10/20/2025).

97. M. Long [et al.] / Learning transferable features with deep adaptation networks

// International conference on machine learning. — 2015. — P. 97-105.

98. B. Sun, J. Feng, K. Saenko / Return of frustratingly easy domain adaptation

// Proceedings of the AAAI conference on artificial intelligence. — 2016. — Vol. 30, Is. 1. — P. 2058-2065.

99. S. Ioffe, C. Szegedy / Batch normalization: Accelerating deep network training by reducing internal covariate shift // International conference on machine learning. — 2015. — P. 448-456.

100. Y. Li [et al.] / Adaptive batch normalization for practical domain adaptation // Pattern Recognition. — 2018. — Vol. 80. — P. 109-117.

101. D. Ulyanov, A. Vedaldi, V. Lempitsky / Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis

// Proceedings of the IEEE conference on computer vision and pattern recognition. — 2017. — P. 6924-6932.

102. G. French, M. Mackiewicz, M. Fisher / Self-ensembling for visual domain adaptation [Electronic Resource]. — 2017. — URL: https://arxiv.org/abs/1706. 05208 (visited on 10/20/2025).

103. Y. Zou [et al.] / Unsupervised domain adaptation for semantic segmentation via class-balanced self-training // Proceedings of the European conference on computer vision (ECCV). — 2018. — P. 289-305.

104. T.-H. Vu [et al.] / Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation // Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. — 2019. — P. 2517-2526.

105. E. Tzeng [et al.] / Adversarial discriminative domain adaptation // Proceedings of the IEEE conference on computer vision and pattern recognition. — 2017. — P. 7167-7176.

106. Y.-H. Tsai [et al.] / Learning to adapt structured output space for semantic segmentation // Proceedings of the IEEE conference on computer vision and pattern recognition. — 2018. —P. 7472-7481.

107. I. Goodfellow [et al.] / Generative adversarial networks // Communications of the ACM. —2020. — Vol. 63, Is. 11.—P. 139-144.

108. J.-Y. Zhu [et al.] / Unpaired image-to-image translation using cycle-consistent adversarial networks // Proceedings of the IEEE international conference on computer vision. — 2017. — P. 2223-2232.

109. Z. Zhang, L. Yang, Y. Zheng / Translating and segmenting multimodal medical volumes with cycle-and shape-consistency generative adversarial network // Proceedings of the IEEE conference on computer vision and pattern Recognition. — 2018. — P. 9242-9251.

110. Y. Yang, S. Soatto / Fda: Fourier domain adaptation for semantic segmentation

// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. — 2020. — P. 4085-4095.

111. I. Zakazov [et al.] / Feather-light Fourier domain adaptation in magnetic resonance imaging // MICCAI Workshop on Domain Adaptation and Representation Transfer. — 2022. — P. 88-97.

112. E. Kondrateva [et al.] / Negligible effect of brain MRI data preprocessing for tumor segmentation // Biomedical Signal Processing and Control. — 2024. — Vol. 96.—P. 106599.

113. C. Ouyang [et al.] / Causality-inspired single-source domain generalization for medical image segmentation // IEEE Transactions on Medical Imaging. — 2022. — Vol. 42, Is. 4. — P. 1095-1106.

114. H. Wang [et al.] / Sam-med3d: towards general-purpose segmentation models for volumetric medical images // European Conference on Computer Vision. — 2024.—P. 51-67.

115. J. Liu [et al.] / Clip-driven universal model for organ segmentation and tumor detection // Proceedings of the IEEE/CVF International Conference on Computer Vision. —2023. —P. 21152-21164.

116. B. Menze [et al.] / Analyzing magnetic resonance imaging data from glioma patients using deep learning // Computerized medical imaging and graphics. — 2021. —Vol. 88.—P. 101828.

117. F. Kofler [et al.] / Brats toolkit: translating brats brain tumor segmentation algorithms into clinical and scientific practice // Frontiers in neuroscience. — 2020.— Vol. 14.—P. 125.

118. F. Isensee [et al.] / nnu-net: Self-adapting framework for u-net-based medical image segmentation [Electronic Resource]. — 2018. — URL: https://arxiv.org/ abs/1809.10486 (visited on 10/20/2025).

119. P. Kickingereder [et al.] / Automated quantitative tumour response assessment of MRI in neuro-oncology with artificial neural networks: a multicentre, retrospective study // The Lancet Oncology. — 2019. — Vol. 20, Is. 5. — P. 728-740.

120. S. Bakas [et al.] / Advancing the cancer genome atlas glioma MRI collections with expert segmentation labels and radiomic features // Scientific data. — 2017.— Vol. 4, Is. 1.—P. 1-13.

121. B. Kompa, J. Snoek, A. L. Beam / Second opinion needed: communicating uncertainty in medical machine learning // NPJ Digital Medicine. — 2021. — Vol. 4, Is. 1.—P.4.

122. J. Yang [et al.] / Generalized out-of-distribution detection: A survey // International Journal of Computer Vision. — 2024. — P. 1-28.

123. D. Hendrycks, K. Gimpel / A baseline for detecting misclassified and out-of-distribution examples in neural networks [Electronic Resource]. — 2016. — URL: https://arxiv.org/abs/1610.02136 (visited on 10/20/2025).

124. S. Basart [et al.] / Scaling out-of-distribution detection for real-world settings [Electronic Resource]. — 2022. — URL: https ://arxiv. org/abs/1911.11132 (visited on 10/29/2025).

125. A. Mahmood, J. Oliva, M. Styner / Multiscale score matching for out-of-distribution detection [Electronic Resource]. — 2020. — URL: https : //arxiv.org/abs/2010.13132 (visited on 10/20/2025).

126. A. G. Pacheco [et al.] / On out-of-distribution detection algorithms with deep neural skin cancer classifiers // Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. — 2020. — P. 732-733.

127. C. Berger [et al.] / Confidence-based out-of-distribution detection: a comparative study and analysis // International Workshop on Uncertainty for Safe Utilization of Machine Learning in Medical Imaging. — 2021. — P. 122-132.

128. T. Cao [et al.] / A benchmark of medical out of distribution detection [Electronic Resource]. — 2020. — URL: https://arxiv.org/abs/2007.04250 (visited on 10/20/2025).

129. G. Litjens [et al.] / A survey on deep learning in medical image analysis // Medical image analysis. — 2017. — Vol. 42. — P. 60-88.

130. D. Karimi, A. Gholipour / Improving calibration and out-of-distribution detection in deep models for medical image segmentation // IEEE transactions on artificial intelligence. — 2022. — Vol. 4, Is. 2. — P. 383-397.

131. D. Zimmerer [et al.] / Medical Out-of-Distribution Analysis Challenge 2022 [Electronic Resource]. — 2022. — URL: https: //doi. org /10. 5281 /zenodo. 6362313 (visited on 10/30/2025).

132. B. Lambert [et al.] / Improving Uncertainty-based Out-of-Distribution Detection for Medical Image Segmentation [Electronic Resource]. — 2022. — URL: https: //arxiv.org/abs/2211.05421 (visited on 10/30/2025).

133. P. Bilic [et al.] / The liver tumor segmentation benchmark (lits) // Medical Image Analysis. — 2023. — Vol. 84. — P. 102680.

134. M. Hssayeni [et al.] / Computed tomography images for intracranial hemorrhage detection and segmentation // Intracranial hemorrhage segmentation using a deep convolutional model. Data. — 2020. — Vol. 5, Is. 1. — P. 14.

135. J. Shapey [et al.] / Segmentation of vestibular schwannoma from MRI, an open annotated dataset and baseline algorithm // Scientific Data. — 2021. — Vol. 8, Is. 1. —P. 286.

136. R. Dorent [et al.] / Cross-Modality Domain Adaptation Challenge 2022 (crossMoDA) [Electronic Resource]. — 2022. — URL: https://doi.org/10. 5281/zenodo.6504722 (visited on 10/30/2025).

137. S. R. van der Voort [et al.] / The Erasmus Glioma Database (EGD): Structural MRI scans, WHO 2016 subtypes, and segmentations of 774 patients with glioma // Data in brief. — 2021. — Vol. 37. — P. 107191.

138. F. Pérez-García, R. Sparks, S. Ourselin / TorchIO: a Python library for efficient loading, preprocessing, augmentation and patch-based sampling of medical images in deep learning // Computer Methods and Programs in Biomedicine. — 2021.—P. 106236.

139. J. Yang [et al.] / Openood: Benchmarking generalized out-of-distribution detection // Advances in Neural Information Processing Systems. — 2022. — Vol. 35.—P. 32598-32611.

140. A. Jungo, M. Reyes / Assessing reliability and challenges of uncertainty estimations for medical image segmentation // International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2019. — P. 48-56.

141. D. Zimmerer [et al.] / MOOD 2020: A public Benchmark for Out-of-Distribution Detection and Localization on medical Images // IEEE Transactions on Medical Imaging. — 2022. — Vol. 41, Is. 10. — P. 2728-2738.

142. A. Mehrtash [et al.] / Confidence calibration and predictive uncertainty estimation for deep medical image segmentation // IEEE transactions on medical imaging. — 2020. — Vol. 39, Is. 12. — P. 3868-3878.

143. L. Smith, Y. Gal / Understanding measures of uncertainty for adversarial example detection [Electronic Resource]. — 2018. — URL: https://arxiv.org/abs/1803. 08533 (visited on 10/20/2025).

144. B. Lakshminarayanan, A. Pritzel, C. Blundell / Simple and scalable predictive uncertainty estimation using deep ensembles // Advances in neural information processing systems. — 2017. — Vol. 30. — P. 6405-6416.

145. Y. Gal, Z. Ghahramani / Dropout as a bayesian approximation: Representing model uncertainty in deep learning // international conference on machine learning. — 2016. — P. 1050-1059.

146. Y.-C. Hsu [et al.] / Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data // Proceedings of the IEEE/CVF

Conference on Computer Vision and Pattern Recognition. — 2020. — P. 10951-10960.

147. Y. Liang [et al.] / Omni-frequency channel-selection representations for unsupervised anomaly detection // IEEE Transactions on Image Processing. — 2023. — Vol. 32. — P. 4327-4340.

148. F. Meissen [et al.] / On the pitfalls of using the residual error as anomaly score //

International Conference on Medical Imaging with Deep Learning. — 2022. — P. 914-928.

149. J. Cho, I. Kang, J. Park / Self-supervised 3D Out-of-Distribution Detection via Pseudoanomaly Generation // International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2021. — P. 95-103.

150. F. Isensee [et al.] / No new-net // International MICCAI Brainlesion Workshop. — 2018. — P. 234-244.

151. N. Abraham, N. M. Khan / A novel focal tversky loss function with improved attention u-net for lesion segmentation // 2019 IEEE 16th international symposium on biomedical imaging (ISBI2019). — 2019. — P. 683-687.

152. A. Vasiliuk [et al.] / Redesigning out-of-distribution detection on 3d medical images // International workshop on uncertainty for safe utilization of machine learning in medical imaging. — 2023. — P. 126-135.

153. A. Vasiliuk [et al.] / Exploring structure-wise uncertainty for 3d medical image segmentation // International Conference on Medical Imaging and Computer-Aided Diagnosis. — 2022. — P. 15-26.

154. M. Goncharov [et al.] / vox2vec: A Framework for Self-supervised Contrastive Learning of Voxel-Level Representations in Medical Images //

International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2023. — P. 605-614.

155. M. Goncharov [et al.] / Anatomical positional embeddings // International Conference on Medical Image Computing and Computer-Assisted Intervention. — 2024. — P. 68-77.

156. M. Goncharov [et al.] / Screener: Self-supervised Pathology Segmentation Model for 3D Medical Images [Electronic Resource]. — 2025. — URL: https: //arxiv.org/abs/2502.08321 (visited on 10/20/2025).

Author's publications on the dissertation subject

157. B. Shirokikh [et al.] / Systematic clinical evaluation of a deep learning method for medical image segmentation: radiosurgery application // IEEE Journal of Biomedical and Health Informatics. — 2022. — Vol. 26, Is. 7. — P. 3037-3046.

158. M. Goncharov [et al.] / CT-Based COVID-19 triage: Deep multitask learning improves joint identification and severity quantification // Medical image analysis. — 2021. — Vol. 71. — P. 102054.

159. B. Shirokikh [et al.] / First U-Net layers contain more domain specific information than the last ones // MICCAI Workshop on Domain Adaptation and Representation Transfer. — 2020. — P. 117-126.

160. I. Zakazov [et al.] / Anatomy of domain shift impact on U-Net layers in MRI segmentation // International Conference on Medical Image Computing and Computer-Assisted Intervention. —2021. —P. 211-220.

161. T. Saparov [et al.] / Zero-shot domain adaptation in ct segmentation by filtered back projection augmentation // MICCAI Workshop on Deep Generative Models. —2021. — P. 243-250.

162. S. Shimovolos [et al.] / Adaptation to CT Reconstruction Kernels by Enforcing Cross-Domain Feature Maps Consistency // Journal of Imaging. — 2022. — Vol. 8, Is. 9. — P. 234.

163. B. Shirokikh [et al.] / M3DA: Benchmark for Unsupervised Domain Adaptation in 3D Medical Image Segmentation [Electronic Resource]. — 2025. — URL: https://arxiv.org/abs/2502.17029 (visited on 9/21/2025).

164. A. Vasiliuk [et al.] / Limitations of out-of-distribution detection in 3d medical image segmentation // Journal of Imaging. — 2023. — Vol. 9, Is. 9. — P. 191.

165. S. Zolotova [et al.] / Burdenko's Glioblastoma Progression Dataset (Burdenko-GBM-Progression) (Version 1) [Data set] [Electronic Resource]. — 2023. — URL: https://doi.org/10.7937/E1QP-D183 (visited on 10/30/2025).

166. М. Г. Беляев, М. Писов, Б. Н. Широких / Автоматическая классификация и сегментация очагов СОУГО-19 на изображениях компьютерной томографии грудной клетки // Свидетельство о гос. регистрации программы для ЭВМ. —N0. 2021612647, 2/20/2021 (ВД).