Comparative analysis of modifications of U-Net neuronal network architectures in medical image segmentation

Full Text

Introduction

Image processing using artificial-intelligence (AI)-based software plays a central role in modern medical diagnosis. Advancements in computational technology and machine learning algorithms have considerably expanded the capabilities of image analysis in recent decades. Comprehensive clinical decision support systems, including autonomous models, have replaced the previous generation of simple classification frameworks.

Medical image processing initially relied on basic imaging modalities such as radiography and mammography. These modalities have since evolved, and computed tomography (CT) and magnetic resonance imaging (MRI) data are now processed with high efficiency. In the context of diagnostic radiology, AI-based software is applied to a range of tasks, including data visualization, segmentation, recording, classification, and interpretation.

Among these, medical image segmentation remains among the most challenging tasks, as it involves identifying clusters of pixels that correspond to specific image objects, particularly in CT and MRI scans. Deep learning algorithms have demonstrated promising performance in segmenting abnormal regions (selecting target regions) and subsequently classifying them. These algorithms notably outperform conventional approaches in both processing accuracy and speed [1]. Various neural network architectures have been employed for segmentation tasks. These models differ in structural characteristics, including the number of layers, neurons per layer, activation functions, and optimization algorithms. Among these architectures, frameworks such as U-Net, V-Net, DenseNet, and Mask R-CNN have demonstrated strong performance in segmentation tasks [2–6].

Since its introduction in 2015, the U-Net segmentation network has become a standard tool in biomedical image processing. Nevertheless, the basic U-Net architecture continues to demonstrate strong performance in analyzing medical images for detecting organ abnormalities, such as those seen in kidney CT scans and lung changes associated with COVID-19 or obstructive pulmonary disease [7–9]. The U-Net3D architecture extends the conventional U-Net by replacing two-dimensional (2D) convolutions with 3D convolutions [10]. It is employed for the segmentation of 3D medical images. For instance, Pantovic et al. used U-Net3D to analyze CT scans of a brain containing neural implants to identify surgical sites for epileptogenic zone removal [11]. Han et al. used the same architecture to segment liver MRI scans and delineate both the contours and internal structures of the liver [12].

The standard architecture of the U-Net neural network consists of two primary components: the encoder and decoder. The encoder compresses the input data and extracts the most relevant features for subsequent recognition. Meanwhile, the decoder reconstructs a segmented image from the compressed data generated by the encoder. Since 2015, numerous modifications to the standard U-Net architecture (referred to as the U-architecture) (Fig. 1) have been developed to enhance its accuracy, speed, and robustness. These modifications can be grouped into four main categories: (1) modifying the encoder and decoder while preserving the overall network structure; (2) combining multiple U-architecture models through ensembling; (3) integrating additional architectural components, such as attention blocks; and (4) incorporating supplementary features into the model.

 

Fig. 1. Classic U-Net architecture proposed in 2015 and the main categories of its modification methods.

 

These modifications have also been applied to address image segmentation challenges arising during semi-supervised learning or when training data are limited (Fig. 2). The limited-training-data scenario can be further categorized into cases involving small and extremely small datasets. Specifically, when training on small datasets, transfer learning and fine-tuning are typically applied to networks pretrained on more diverse datasets.

 

Fig. 2. Segmentation tasks categorized by the availability and type of training data.

 

Meanwhile, when training on extremely small datasets (few-shot learning), pretraining is inadequate; such cases generally require original architectures and data models.

This review explores the application of U-Net architecture modifications in medical image processing. Section 1 outlines the main modification strategies for the U-architecture, including (1) changing the encoder and decoder internally, (2) integrating additional architectural components such as attention blocks, and (3) altering the network’s learning process. Section 2 explores how these modifications can be applied to address specific challenges in medical image segmentation. The conclusion in Section 3 summarizes the key findings of the review.

Data search methodology

The authors conducted a literature search using the Web of Science, Scopus, and PubMed databases, covering publications from 2018 to 2024. The search results were comparable across databases and reflected the primary trends in U-Net architecture modification methods. The search keywords included U-Net, medical images, and modification. The initial search returned approximately 5,000 sources. This was subsequently refined using additional terms, including attention, few-shot, unsupervised, semi-supervised, ensemble, stack, additional features, metadata, and DICOM data.

The selected publications were reviewed with a focus on the use of specific architectures for medical image processing. The inclusion criteria were as follows:

• Quality of result validation (e.g., comparison with other architectures, use of established evaluation metrics, and study completeness);
• Originality of the architectural modification in relation to its intended application;
• Specificity of the task (e.g., type of abnormality detected or organ segmented);
• Use of open datasets.

The U-Net architecture has substantially impacted medical image segmentation owing to its effectiveness. Originally proposed by Ronneberger et al. [2], U-Net has since evolved into several notable variants, including U-Net++ [13], Attention U-Net [5], 3D U-Net [10], EU-Net [14], NAS-U-Net [15], U-Net 3+ [16], and SwinAttU-Net [17]. Appendix 1 provides an overview of key studies on U-Net modification methods and segmentation accuracy evaluations, as well as datasets used for testing. It also includes studies wherein U-Net was applied to address specific segmentation challenges. The following abbreviations are used for performance metrics: DC, Dice coefficient; IoU, intersection over union; OA, overall accuracy [18, 19].

U-net architecture modifications

Internal encoder and decoder modifications

This section discusses structural elements that are altered by internal modifications to the encoder and decoder of the U-architecture.

Encoder and decoder convolution blocks. To process spinal cord images (Verse2019 and Verse2020 datasets), Xu et al. replaced convolution layers with linear layers in the encoder and with octave convolutions in the decoder. Octave convolutions combine standard convolution blocks with pooling operations to extract frequency-based data [57]. Ayalew et al. reduced the number of convolution channels and incorporated batch normalization into the original U-architecture to detect liver tumors in CT scans [58]. This modification improved network accuracy on datasets with considerable class imbalance. Guan et al. proposed an architecture with modified convolution blocks wherein the outputs of each layer were concatenated and jointly processed to minimize distortion in photoacoustic images, such as brain scans [59].

Connections between encoder and decoder blocks. Özcan et al. used a U-Net variant to identify tumor regions in liver CT scans. In this variant, connections between encoder and decoder blocks passed through an inception block composed of convolutions with different kernel sizes, whose outputs were concatenated. In other studies, these connections passed through a pyramid of pooling layers (consisting of multiple pooling layers with different kernel sizes applied to the same data) [61]. This approach was used to accelerate the segmentation of liver ultrasound images.

Encoder or decoder regularization blocks. Omarov et al. applied a modified U-Net architecture to detect brain regions affected by ischemic stroke on CT scans. In this architecture, dropout and L2 regularization layers were incorporated into the decoder [62].

Ensembling U-Net architectures. A concatenated ensemble of U-Net networks trained on ImageNet images converted to sinograms was used to reconstruct CT images from projection data obtained by rotating an object [63]. In another example, an ensemble of two U-Net3D networks pretrained on the LiTS dataset was applied for detecting liver tumors in 3D CT scans [24, 64]. The first network processed low-resolution images (reduced source images), and its segmentation output was passed to the second network. A combined loss function incorporating the DC and cross-entropy was used. In another study, a two-stage U-Net ensemble was developed for liver tumor detection. Here, one network functioned as a post-processing and refinement stage [65].

Koirala et al. used an ensemble of U-Net3D, ONet3D, and SphereNet3D networks to locate brain tumors. Ensembling was achieved by weighing (summing and multiplying by a number reflecting the network’s contribution to the overall result, i.e., its weight) the outputs of all models to determine the most probable class.

Li et al. used an unmodified U-Net architecture to select the optimal model for their application [67].

Overall, existing studies suggest that even minor architectural changes to U-Net can improve its effectiveness in medical imaging tasks.

Modifications using attention mechanisms

This section outlines how previous studies modified the standard U-Net architecture by integrating spatial and channel attention blocks [68]. In one study, a U-Net3D-based variant incorporating efficient channel attention in the encoder blocks was applied to detect COVID-19-related abnormalities in chest CT scans [69]. In another study, a pyramid fusion module was implemented at the lower layer of the U-architecture. In this module, features extracted using neural networks with varying window sizes were concatenated, and the resulting data were processed using a pooling layer with a global mean value. The Tversky loss function was used for optimization [70].

Another study focused on the simultaneous segmentation of multiple organs using CT scans [71]. The proposed U-Net architecture included an attention block that took the outputs of both the encoder and decoder as input. These outputs were concatenated and processed using a 1D convolution operation with ReLU activation sigmoid functions [72].

One study [73] employed a U-Net architecture with spatial multi-scale attention blocks to segment liver tumors in CT scans. These attention blocks were placed at multiple points in the architecture, including within the encoder and decoder, as well as along the connections between them.

Zhang et al. applied pyramid pooling in the lower part of the U-architecture (corresponding to the point of maximum data reduction) and used efficient channel attention blocks on the connections between the encoder and decoder blocks [74]. Another study proposed a U-Net architecture with spatial attention between encoder blocks, incorporating convolutions with multiple receptive fields (Fig. 3). This architecture was trained using the Tversky loss function for breast cancer detection [75].

 

Fig. 3. Spatial attention block positioned between encoder elements [75].

 

Subhan Akbar et al. introduced attention blocks into the connections between the encoder and decoder blocks of the U-architecture. For feature extraction, they also added a positional attention block and a self-attention block to each layer of the decoder [76, 77].

Thus, in U-Net, various attention blocks have been used to capture spatial relationships between image elements at different scales. Notably, these relationships cannot be detected using the basic architecture.

Modifications through the integration of elements from other architectures

A common approach to modifying the U-Net architecture involves incorporating elements from other networks, such as ResNet or transformers. Several variations of this approach have been proposed.

Full modification of the encoder and/or decoder. Xingfei et al. modified the U-Net architecture by replacing the encoder with ResNet50 for segmenting COVID-19-related abnormalities in the lungs [78, 79]. A channel attention block combined with a pyramid pooling module was applied following the encoder. Alternatively, a transformer encoder can be integrated, with its output upsampled via deconvolution for use in different parts of the U-architecture [80].

Modification of encoder and decoder blocks while maintaining the general U-Net architecture. Eskandari et al. focused on segmenting liver structures in CT scans [81]. To account for the considerable variability in liver shape, size, and position, they used a position-determining classifier network in combination with a modified U-Net architecture. This modification replaced standard convolution blocks with ConvLSTM blocks, which were also incorporated into the connections between encoder and decoder blocks [82].

In another study, a hybrid architecture combining efficient transformer blocks with the U-Net architecture was proposed for identifying skin abnormalities in medical images (Fig. 4) [83]. This architecture outperformed the classic U-Net, Attention U-Net, TransU-Net, FAT-Net, and Swin U-Net in terms of DC, sensitivity, specificity, and accuracy on the ISIC 2018 skin lesion dataset.

 

Fig. 4. Architecture integrating transformer blocks into the U-Net framework [81].

 

Ghofrani et al. applied a combination of an unmodified U-Net and transformer blocks to segment polyp images, achieving higher accuracy than U-Net, ResU-Net++, and DoubleU-Net [36, 37, 84–86].

For 3D liver image segmentation, U-Net was combined with Swin Transformer, BTSwin Transformer, and DenseNet components [87–89].

In summary, similar to attention-based modifications, integrating elements from other architectures enhances image processing quality by identifying subtle relationships between image regions. Transformer blocks that employ self-attention mechanisms to extract latent features are frequently used in this context.

Introducing additional features into u-net

Researchers often use metadata from DICOM files as supplementary features in medical image analysis. These data are typically tabulated and include both continuous and categorical variables. The metadata are often input into a separate network, which may be trained either jointly with or independently from the main segmentation model. This supplementary information is generally incorporated into the base network using attention mechanisms. For instance, in a previous study on spinal tumor segmentation, metadata were integrated into a U-Net-based segmentation model. Each block included a linear transformation block applied to the output of the preceding convolutional layer [90]. The U-Net-based generator computed transformation parameters (shift and scale) after receiving metadata related to the segmented image. In another study, Du et al. proposed a channel attention mechanism wherein metadata were used to train the 3D-RADNet network to detect image slices containing the target organ (liver) [91]. Slices selected using metadata were processed by a U-Net-based segmentation model. In kidney tumor segmentation, channel attention has been used to incorporate metadata into the network, allowing the data to serve as the outputs of U-Net blocks [92]. After the final convolutional layer of each block, both image data and metadata are passed to a layer where the metadata are input into a multi-layer perceptron (MLP) with a sigmoid activation function. The MLP outputs are then multiplied, on a per-channel basis, with the image data from the preceding convolutional layer.

In addition to metadata, other sources of auxiliary information have been used to enhance U-Net models:

• A two-branch architecture based on a convolutional network [93];
• CNNFormer for liver segmentation, which accounts for both intra-slice spatial relationships and inter-slice hierarchical structures [94];
• Additional features, such as spine, lung, and skin segmentation results obtained using the Python library Body Navigation [95].

These data have been concatenated with the input images to enhance the localization of the target organ. This approach has been applied to liver CT segmentation using both U-Net and U-Net3D architectures, depending on whether individual slices or entire scans were processed.

 

Fig. 5. Combined U-Net and transformer-based architecture [83].

 

Many modifications to U-Net training involve the iterative reuse of features. For example, Ernst et al. focused on reconstructing CT images from sinograms [96]. They employed a combination of U-Net3D and Primal-Dual networks with iterative learning, where the output at each step was combined with the results of the previous iteration. Another study proposed a method to improve segmentation accuracy by reusing features extracted during learning [97]. RecycleNet, an architecture derived from U-Net, comprises three main blocks:

• I: input data block;
• R: latent feature reuse block;
• O: outcome block (Fig. 6).

 

Fig. 6. Structural blocks of the U-Net architecture [97].

 

The feature reuse algorithm is illustrated in Fig. 6. First, the number of iterations to be used for decision-making is randomly selected from a predefined range. The features extracted in the previous iteration are normalized and added to those from the current iteration, incorporating spatial embedding. After completing the selected number of iterations, the network generates the final output. In a previous study, RecycleNet was experimentally evaluated on the KiTS 2019 (kidney cancer), LiTS, BTCV, AMOS (multi-organ segmentation), and CHAOS (MRI) datasets [23, 24, 33, 40]. The proposed architecture was compared with a DC-optimized variant of nnU-Net and the DRU network [98]. RecycleNet outperformed the compared architectures on all evaluated datasets.

Thus, incorporating additional features can improve the accuracy of image processing using U-Net. Such supplementary data often reveal patterns that are not present or are only weakly expressed in the image itself.

Addressing specific segmentation challenges using the u-net architecture

Transfer learning and fine-tuning of U-Net

In medical image processing, available training datasets are often small and structurally complex. This limitation arises from the difficulties encountered during data labeling and restrictions imposed by privacy agreements. A common approach in such cases is to employ pretrained models and fine-tune them on the available datasets.

Heker et al. investigated liver tumor segmentation using a small dataset of CT scans [99]. To this end, they first trained the U-Net architecture on the LiTS dataset and applied a hierarchical freezing strategy to its encoder weights. Initially, the encoder weights were frozen, meaning they were not updated during training. The rest of the network was trained for a set number of iterations. Afterward, the frozen encoder weights were gradually unfrozen and fine-tuned one by one.

Several researchers employed a U-Net architecture with a ResNet32-based encoder, initially pretrained on ImageNet and subsequently fine-tuned using optical coherence tomography images [100]. Meanwhile, others have explored fine-tuning techniques for U-Net and U-Net3D in the segmentation of various organs and diseases, including approaches involving a variable number of trainable layers [101, 102].

Moreover, transfer learning with U-Net and EfficientNet architectures—both originally developed for 2D image segmentation—has been applied to facilitate data transfer during 3D image processing [103, 104]. The authors of the aforementioned paper proposed two approaches: 1) increasing the sampling rate of 2D weights in the corresponding blocks of 3D architectures and 2) obtaining plane projections of 3D data and subsequently processing them using a network trained on 2D data (Fig. 7).

 

Fig. 7. Ratios of labeled and unlabeled data in network training and testing: (a) semi-supervised learning (SSL), (b) unsupervised domain adaptation (UDA), and (c) semi-supervised domain generalization (SemiDG) [106].

 

Another approach to training involves using U-Net for post-processing image segmentation results. Hong et al. applied this strategy for liver segmentation in CT scans. In their proposed modification, U-Net’s segmentation output underwent post-processing through the optimization of an energy functional. This functional included two components: one for contour delineation in an image and another for optimizing voxel class labels within the evaluated region.

The effectiveness of fine-tuning and transfer learning strategies strongly depends on the datasets used during pretraining. The closer the training and target datasets are in terms of the types of objects assessed, the more effective fine-tuning and transfer learning become. However, achieving this similarity is not always feasible, particularly for specialized tasks. Large datasets are often unavailable—especially for 3D data. A promising alternative is to fine-tune using simpler, lower-dimensional data, which are generally easier to collect in sufficient quantities.

Semi-supervised learning methods

The shortage of sufficient training data for complex architectures is often due to the lack of expert annotation of raw data—a task that requires substantial domain-specific knowledge and expertise. To address this limitation, various training strategies based on the U-Net architecture have been developed to leverage unlabeled data and semi-supervised learning approaches.

Wang et al. explored the training of segmentation networks for 3D organ models using semi-supervised learning techniques [106]. They developed a framework capable of handling different proportions of labeled and unlabeled data during both training and testing phases (See Fig. 7):

• Fig. 7(a): labeled and unlabeled data, as well as testing data, are of the same type (testing data indicated with a dotted line);
• Fig. 7(b): labeled and unlabeled data are of different types;
• Fig. 7(c): the training set contains labeled and unlabeled data of different types, while the testing data are entirely distinct from both.

The resulting framework consists of two main components (Fig. 8): an aggregation block and a decoupling block. The aggregation block includes the encoder of the proposed Diffusion VNet, which performs image segmentation for type 1 relationships. The decoupling block contains three VNet decoders, each responsible for generating class labels of a specific type. The first decoder produces labels that are unbiased with respect to the type of labeled data, using a loss function that combines cross-entropy and DC. These labels are then used to generate re-weighted class labels, where the weights are applied in a loss function consisting of the sum of DCs across all labeled data classes. This weighting strategy enhances the training effectiveness for classes that perform poorly. The second decoder generates class pseudo-labels for unlabeled data, which are subsequently used to train the third decoder in an unsupervised manner.

 

Fig. 8. A&D framework [106].

 

In a previous study, the above framework was trained using the LASeg (brain MRI), Synapse (various organs), MMWHS, and M&Ms (heart) datasets [47–50]. Its performance was evaluated against that of UA-MT, LMISA-3D, vMFNet, SS-Net, and other architectures using metrics such as DC, Jaccard index, and HD95. In several cases, the framework demonstrated performance that was either superior to or comparable with that of specialized architectures.

Wang et al. investigated trained network adaptation for segmenting a small target dataset focused on polyp detection [107]. The study evaluated a scenario wherein the target dataset consisted of images similar to those used for network training but lacked labels. Two techniques were applied for training: contrastive learning and pseudo-labeling with calibration.

In the contrastive learning phase, unlabeled images were labeled as either positive (consistent with a given image) or negative. Images obtained through augmentation were treated as positive, while others were treated as negative. A network trained on a different dataset generated pseudo-masks for the target dataset. These predicted masks were then used to calculate entropy and determine class centers within the target scans.

To improve the reliability of the generated pseudo-masks, a per-pixel calibration block was introduced. This block incorporated previous predictions to refine the mask quality. To evaluate the effectiveness of the proposed method in polyp segmentation, experiments were conducted using the ClinicDB, ETIS-LARIB, and Kvasir-SEG datasets. The proposed architecture was compared with other networks employing techniques such as bidirectional learning (BDL), Fourier domain adaptation, historical contrastive learning, and denoised pseudo-labeling. The proposed architecture outperformed these alternatives in terms of DC and IoU variations.

Wang et al. also proposed a method for segmenting human organ images, including those captured during surgery, using semi-labeled datasets.

For unlabeled data processing, a dual-network configuration was used (Fig. 9), in which two networks with identical architectures received the same image input. Although the networks were initialized differently, aggregating their outputs enabled more accurate predictions than either network could achieve independently. To avoid distortion when assigning pseudo-labels to unlabeled data in cases where the training dataset exhibited heterogeneous class representations, individual class distributions were reconstructed rather than relying on the overall data distribution.

 

Fig. 9. Dual-network architecture trained on datasets with heterogeneous class representation [108].

 

To align individual class densities, an exponential moving average transformation was applied to class alignment matrices of both labeled and unlabeled data. The effectiveness of the proposed method was evaluated using the CaDIS (surgical images), LGE-MRI, and ACDC (heart disease) datasets. Its performance was compared with that of the URPC, UAMT, CLD, and CPS architectures using the DC, Jaccard index, and additional metrics. The proposed method outperformed all of these architectures across the evaluated parameters.

Thus, a properly selected architecture enables the effective use of unlabeled data in training U-Net-based models, even in the presence of class imbalance.

U-Net training using extremely small sets of real-world data

Developing AI-based software for specific medical tasks is hindered by the challenge of assembling a sufficiently large training dataset [109]. In many cases, dedicated tools are required to process and structure text-based protocols [110–112]. Combined with the high cost of data annotation, these challenges frequently force developers to work with limited amounts of labeled data for machine learning. Consequently, few-shot learning has become a widely adopted approach in medical image processing.

A study investigated the use of CT and positron emission tomography scans for lung cancer detection [113]. A standard U-Net architecture without modifications was trained using data augmentation, with additional data introduced during both training and testing phases based on feedback from an expert evaluating the model’s performance. A similar approach was later applied to COVID-19 data [114]. In another study, the encoder of the U-Net architecture was modified using a Siamese-Net-type structure to enhance segmentation quality. A second encoder branch was introduced; it received the image multiplied by its corresponding mask (segment). The weights from this branch were then combined with those of the primary encoder branch, which processed the original, unmodified image [115].

In the context of medical imaging, this approach is more frequently applied to architectures other than U-Net, which may be due to the network’s size and the number of neurons it contains.

Conclusion

The classic U-Net architecture has proven highly effective for medical image segmentation, which explains its widespread use and the ongoing development of various performance-enhancing modifications. These modifications are designed to improve the interpretation of available data and to pool features obtained during pretraining on diverse datasets, including those that are unlabeled. U-Net modifications can also be categorized according to their intended tasks—such as segmentation or the detection of affected tissues—as well as by the types of datasets used, particularly those representing specific diseases. Additionally, the diagnostic accuracy of U-Net-based solutions can be further enhanced by incorporating supplementary training features derived from text, tabular data, or mathematical models.

U-Net architectures are applied across a wide range of medical image segmentation tasks, which vary in both problem formulation and data type (various types of images and diseases). Each task presents its own unique challenges, making it difficult to define a single, universally effective architecture or even a universally applicable class of models. However, among the approaches assessed, U-Net modifications incorporating elements from other architectures demonstrate the strongest performance. These hybrid models are effective for standard image segmentation tasks—particularly when integrating transformer blocks—as well as for situations where training data are limited, such as through pretraining with networks of lower dimensionality than the target data. The integration of additional features into neural network architectures also shows promise. Similarly, the application of physics-informed neural networks, which incorporate information on object models or image structure, is another promising direction [116–119].

Additional information

Appendix 1. Ways to modify the U-Net architecture. doi: 10.17816/DD629866-4224037

Funding source. This article was prepared by a group of authors as a part of the research and development effort titled «Development of a platform for improving the quality of AI services for clinical diagnostics» (USIS No.: 123031400006-0) in accordance with the Order No. 1196 dated December 21, 2022 «On approval of state assignments funded by means of allocations from the budget of the city of Moscow to the state budgetary (autonomous) institutions subordinate to the Moscow Health Care Department, for 2023 and the planned period of 2024 and 2025» issued by the Moscow Health Care Department. The research was carried out using the infrastructure of the federal state budgetary educational institution of higher education «MIREA – Russian Technological University» within the framework of additional agreement No. 1 dated November 24, 2023 to the cooperation agreement No. 1 dated 07.07.2022, (Moscow).

Competing interests. The authors declare that they have no competing interests. Figures 1 and 2 are original and made by the authors. Figures 3-9 are distributed under the CC BY 4.0 license and are presented in this work unchanged with reference to the original works where they were first presented.

Authors’ contribution. All authors made a substantial contribution to the conception of the work, acquisition, analysis, interpretation of data for the work, drafting and revising the work, final approval of the version to be published and agree to be accountable for all aspects of the work. A.M. Dostovalova — collection and processing of materials, writing the text of the article; A.K. Gorshenin — problem statement, analysis and systematization of approaches, conceptualization, writing the text of the article; Ju.V. Starichkova, K.M. Arzamasov — concept of the work, writing the text of the article.

About the authors

Anastasia M. Dostovalova

MIREA — Russian Technological University; Federal Research Center Computer Science and Control of the Russian Academy of Sciences

Author for correspondence.
Email: adostovalova@frccsc.ru
ORCID iD: 0009-0004-9420-4182
SPIN-code: 3784-0791
Russian Federation, Moscow; Moscow

Andrey K. Gorshenin

MIREA — Russian Technological University; Federal Research Center Computer Science and Control of the Russian Academy of Sciences

Email: agorshenin@frccsc.ru
ORCID iD: 0000-0001-8129-8985
SPIN-code: 1512-3425

Dr. Sci. (Physics and Mathematics), Assistant Professor

Russian Federation, Moscow; Moscow

Julia V. Starichkova

MIREA — Russian Technological University

Email: starichkova@mirea.ru
ORCID iD: 0000-0003-1804-9761
SPIN-code: 3001-6791

Cand. Sci. (Engineering), Assistant Professor

Russian Federation, Moscow

Kirill M. Arzamasov

MIREA — Russian Technological University; Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies

Email: ArzamasovKM@zdrav.mos.ru
ORCID iD: 0000-0001-7786-0349
SPIN-code: 3160-8062

MD, Cand. Sci. (Medicine), Head of Medical Informatics, Radiomics and Radiogenomics Department

Russian Federation, Moscow; Moscow

References

Supplementary files