Integrating AI for Human-Centric Breast Cancer Diagnostics: A Multi-Scale and Multi-View Swin Transformer Framework

arXiv:2503.13309 · eess.IV, cs.AI, cs.CV · Submitted 2025-03-17 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "Integrating AI for Human-Centric Breast Cancer Diagnostics".

Dev: Here's the gist of the paper. Integrating AI for Human-Centric Breast Cancer Diagnostics: A Multi-Scale and Multi-View Swin Transformer Framework proposes a hybrid framework to enhance breast cancer diagnosis by addressing reliance on detailed tumor annotations and susceptibility to missing views during test time The proposed MSMV-Swin framework is designed to be a decision-support tool, helping radiologists analyze multi-view mammograms more effectively.

Rosa: Framework Concept The core of the paper introduces the hybrid, multi-scale and multi-view Swin Transformer-based framework called MSMV-Swin to enhance diagnostic robustness and accuracy.

Dev: This framework leverages the Segment Anything Model (SAM) to isolate the breast lobe, reducing background noise and enabling comprehensive feature extraction.

Rosa: The multi-scale nature accounts for tumor-specific regions and the spatial characteristics of tissues surrounding the tumor, capturing both localized and contextual information.

Paper summary: Rosa: So we're looking at this paper called "Integrating AI for Human-Centric Breast Cancer Diagnostics: A Multi-Scale and Multi-View Swin Transformer Framework," and it’s about trying to make breast cancer diagnosis better by using AI in a way that actually helps the doctor.

Dev: That sounds like it tackles some real practical problems right away, Rosa. What's the main argument they're making with this paper?

Rosa: The core idea is that current systems often rely too much on very detailed drawings or annotations of tumors, and they struggle when you don't have all the necessary views during a test. This framework proposes a hybrid approach using a multi-scale and multi-view Swin Transformer to fix those issues.

Dev: So it’s addressing the problem of needing perfect tumor outlines and dealing with incomplete information in mammograms, which is something radiologists face every day during testing, right?

Rosa: Exactly. The paper claims this new MSMV-Swin framework acts more like a decision support tool for the radiologist to analyze those multi-view mammograms more effectively. It’s about making the diagnosis process more robust against missing views.

Taro: When you say "robust against missing views," are you talking about how the system handles it if one of those necessary images just doesn't arrive? Does it try to guess what's missing?

Dev: Not guessing, Taro. It seems they use a feature zero padding strategy within the framework itself to provide robustness against those missing views without having to create synthetic data or anything artificial.

Rosa: That’s a key point there. They use the Segment Anything Model, or SAM, at the beginning to isolate the breast lobe from all that background noise so they can start extracting features properly for both scales and views.

Taro: Isolating the lobe with SAM sounds like a good first step to clean up the input data before feeding it into such a complex transformer architecture. What about how they handle the different scales?

Dev: They use two distinct paths for feature extraction, one for segmented images and one for cropped images, which means they are processing features from both those different sizes of inputs separately.

Rosa: That separation is what’s interesting because it lets them capture both the localized tumor-specific regions and the broader spatial characteristics of the tissues surrounding that tumor at once. It’s about getting both the fine detail and the context.

Taro: And then they use a dual Swin Transformer architecture for each scale, which is designed to help it capture those long-range dependencies through a shifted window strategy, right?

Dev: That shifted window strategy is essential because it lets the transformer process images hierarchically and efficiently, which helps integrate local and global contextual features together effectively. It’s how they get that multi-view information flowing early in the network.

Paper summary: Rosa: And to combine all those extracted features, they use two different feature fusion strategies: one involves max pooling across concatenated features, and the other uses a convolutional block for refinement after concatenation into a one thousand twenty-four by four array.

Taro: So, one strategy focuses on selecting only the most relevant features with the highest values through that max pooling, while the other refines things through a custom convolution block that has a small kernel size of three by three.

Dev: That convolutional block then includes batch normalization and ReLU activation before another pooling step, which is how they refine those features into a unified feature map for the final prediction. It’s all about getting that best representation.

Rosa: Experimentally, they tested this MSMV-Swin framework on the CBIS-DDSM dataset, and the results showed superior performance compared to state of the art models like Reference eleven nineteen twenty and twenty-one across both views and missing data scenarios.

Taro: So they found that even when there are missing views, the performance drop is negligible compared to cases where all views are present. That’s a strong indicator for real-world reliability.

Dev: It suggests that this framework can provide a much more stable diagnostic output even in messy, incomplete clinical situations without needing to rely on synthetic data generation. That addresses one of the major hurdles we talked about earlier.

Rosa: Overall, the paper shows that by combining multi-scale input processing with this hybrid fusion structure, they successfully integrate those localized and contextual tissue characteristics around tumors from both CC and MLO views.

Taro: It feels like they’ve built something that accounts for both the specific pathology and the surrounding environment simultaneously. That's a big step forward for how we model these complex biological images.

Dev: And what this means practically is that this system could significantly improve how radiologists make decisions in real clinical settings because it handles data imperfections much better than previous methods.

Rosa: So, to wrap up, the MSMV-Swin framework, as detailed in "Integrating AI for Human-Centric Breast Cancer Diagnostics: A Multi-Scale and Multi-View Swin Transformer Framework," offers a way to build a more resilient diagnostic tool using multi-scale processing and dual feature fusion techniques.

Taro: It moves beyond just looking at one view or one annotation by incorporating the whole context across different sizes and views of the image.

Dev: The implication is that we can start building diagnostic support systems that are less fragile when the input data isn't perfectly clean or complete during live testing loops.

Rosa: So, this work points toward a future where AI integration in diagnosis focuses on making the tool usable and reliable under imperfect conditions, rather than just aiming for perfect accuracy on ideal data sets.

Conclusion: Rosa: So, to wrap up, this paper introduces a framework called MSMV-Swin that uses multi-scale processing and dual feature fusion to improve breast cancer diagnosis by handling missing views better than before.

Dev: Yeah, it’s about taking those complex mammogram images and using this Swin Transformer thing to get a more reliable decision support tool for radiologists.

Rosa: The authors are trying to move away from needing perfect annotations and instead focus on capturing both the small tumor details and the wider tissue context across different image views.

Dev: It’s designed specifically so that even if one of the necessary mammograms is missing during a test, the system can still give a decent result because of how it fuses its features.

Rosa: They found that this approach gives better accuracy compared to other models on datasets like CBIS-DDSM, even in those incomplete scenarios.

Dev: So it’s about making the diagnostic pipeline more resilient when the real world throws curveballs at you with missing data.

Rosa: It means we can start building tools that are more trustworthy for doctors when they aren't getting perfectly clean input every single time.

Dev: This moves the focus from just getting a high score on perfect images to having a system that works reliably in messy, real-world clinical settings.

Rosa: And it opens up a path for AI to be genuinely human-centric by supporting the doctor more effectively when they’re under pressure or have incomplete information.

Dev: It sets the stage for thinking about how these kinds of robust systems could integrate into live diagnostic workflows instead of just sitting in a lab environment.

Farnoush Bayatmakou, Reza Taleei, Milad Amir Toutounchian, Arash Mohammadi

Concordia Institute for Information Systems Engineering, Concordia University · Thomas Jefferson University Hospital, Thomas Jefferson University Hospital, Philadelphia Pennsylvania, USA · College of Computing & Informatics, Drexel University, Philadelphia Pennsylvania, USA

eess.IV, cs.AI, cs.CV

Submitted: 2025-03-17

Updated: 2026-10-30

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: Integrating AI for Human-Centric Breast Cancer Diagnostics: A Multi-Scale and Multi-View Swin Transformer Framework proposes a hybrid framework to enhance breast cancer diagnosis by addressing…

Key concepts

MSMV-Swin Framework
A hybrid framework that combines multi-scale processing with multi-view analysis using the Swin Transformer. It uses the Segment Anything Model (SAM) to isolate breast tissue, capturing both fine tumor details and broader contextual information from different image views.
Multi-Scale Processing
This involves handling mammogram data at different resolutions. The framework processes inputs in two scales: one for localized tumor regions and another for the surrounding tissue characteristics, ensuring that both specific features and general spatial context are captured effectively.
Feature Fusion Strategies
The model combines information from different processing streams using two methods: Max-Pooling Fusion, which selects the most important features, and Convolution-based Fusion. This combination helps create a unified feature map that is robust even when some image views are missing.

Terminology

Summary

Integrating AI for Human-Centric Breast Cancer Diagnostics: A Multi-Scale and Multi-View Swin Transformer Framework proposes a hybrid framework to enhance breast cancer diagnosis by addressing reliance on detailed tumor annotations and susceptibility to missing views during test time The proposed MSMV-Swin framework is designed to be a decision-support tool, helping radiologists analyze multi-view mammograms more effectively. p. 4

Where the paper says it

CONCLUSION In this paper, we introduced the MSMV-Swin framework, a hybrid, multi-scale and multi-view Swin Transformer-based architecture aimed at addressing two key challenges in breast cancer diagnosis through mammography, i. , (i) Reliance on detailed tumor annotations and (ii) Susceptibility to missing views during test time.

Framework Concept

The core of the paper introduces the hybrid, multi-scale and multi-view Swin Transformer-based framework called MSMV-Swin to enhance diagnostic robustness and accuracy. This framework leverages the Segment Anything Model (SAM) to isolate the breast lobe, reducing background noise and enabling comprehensive feature extraction. p. 0

Where the paper says it

The proposed MSMV-Swin framework is designed to be a decision-support tool, helping radiologists analyze multi-view mammograms more effectively. More specifically, the MSMV-Swin framework leverages the Segment Anything Model (SAM) to isolate the breast lobe, reducing background noise and enabling comprehensive feature extraction.

The multi-scale nature accounts for tumor-specific regions and the spatial characteristics of tissues surrounding the tumor, capturing both localized and contextual information. p. 0
Where the paper says it

More specifically, the MSMV-Swin framework leverages the Segment Anything Model (SAM) to isolate the breast lobe, reducing background noise and enabling comprehensive feature extraction. The multi-scale nature of the proposed MSMV-Swin framework accounts for tumor-specific regions and the spatial characteristics of tissues surrounding the tumor, capturing both localized and contextual information.

A hybrid fusion structure is designed to provide robustness against missing views through a feature zero padding strategy. p. 0
Where the paper says it

Integrating contextual and localized data ensures that MSMV-Swin’s outputs align with how radiologists interpret mammograms, fostering better human-AI interaction and trust. A hybrid fusion structure is then designed to provide robustness against missing views, a common occurrence in clinical practice when only a single mammogram view is available.

Model Architecture and Feature Extraction

The MSMV-Swin framework leverages a dual Swin Transformer architecture to process distinct scales of mammographic image inputs—specifically, cropped and segmented variants. The processing pipeline is designed to handle separate streams for CC and MLO views for each of the two underlying scales. p. 1

Where the paper says it

PROBLEM FORMULATION AND PROPOSED MSMV-SWIN FRAMEWORK The MSMV-Swin framework leverages a dual Swin Transformer architecture to process distinct scales of mammographic image inputs—specifically, cropped and segmented variants. The processing pipeline is designed to handle separate streams for CC and MLO views for each of the two underlying scales.

The feature extraction framework involves several key steps:

  1. Pre-processing utilizes a well-trained SAM to segment original mammograms, removing background areas to isolate the breast lobe. p. 1

    Where the paper says it

    Model Architecture and Training As the pre-processing step, we utilized a well-trained SAM to segment original mammograms. This approach is beneficial in enhancing the learning process of the MSMV-Swin framework by removing the background areas from mammogram images to isolate the breast lobe.

  2. Extracted breast areas are downsampled to (224, 224) pixels to reduce complexity without significant loss of information. p. 1

    Where the paper says it

    The extracted breast lobe is then passed into two parallel paths of the proposed MSMV-Swin framework. More specifically, extracted breast areas after pre-processing are downsampled to (224, 224) pixels to reduce the complexity and memory allocation without significant loss of information.

  3. A pre-trained Swin Transformer is deployed in dual configurations, processing segmented and cropped images independently for each scale. p. 2

    Where the paper says it

    Feature Extraction Framework Pre-trained Swin Transformer is deployed in dual configurations given its capabilities to process images hierarchically and efficiently capture long-range dependencies through a shifted window strategy, which is essential for integrating local and global contextual features effectively. The architecture integrates dual instances of Swin Transformers, each fit to the unique characteristics of the corresponding scale and set to process segmented and cropped images independently.

Feature Fusion Strategies

The framework employs two distinct feature fusion strategies to combine the extracted features into a unified feature map.

  1. Max-Pooling Fusion involves concatenating the four features and then performing a reduction process where maximum values are selected across concatenated features. This step ensures only the most relevant features with the highest values are passed forward. p. 3

    Where the paper says it

    This operation is performed across dimensions with different views (CC and MLO) and segmentation statuses (segmented and cropped), effectively capturing the most critical features. This maxpooling step ensures that only the most relevant features with the highest values post concatenation are passed forward, enhancing the model’s focus on potentially critical diagnostic information.

  2. Convolution-based Fusion concatenates the outputs into a 1024 × 4 array and passes it through a custom-designed convolutional block for refinement. This block consists of a 2D convolutional layer with a kernel size of 3 × 3, followed by batch normalization, ReLU activation, and max-pooling. p. 3

    Where the paper says it

    This combination not only maximizes the feature extraction capabilities of Swin Transformers but also leverages the robust, spatial feature processing of convolutional networks, significantly improving the accuracy and reliability of image classification tasks by ensuring a comprehensive and refined analysis of input data. The incorporated block consists of a 2D convolutional layer with a kernel size of 3 × 3, followed by batch normalization, a ReLu activation function, and a max-pooling layer.

Experimental Validation

Experimental results on the CBIS-DDSM dataset demonstrate the superior performance of MSMV-Swin. The model achieves better accuracy and robustness even when dealing with missing views, showing a negligible performance drop compared to dual-view cases. p. 4

Where the paper says it

Experimental results demonstrate that the MSMV-Swin framework consistently outperforms existing models. Notably, it achieves superior accuracy and robustness even in scenarios where only single-view mammograms are available, with a negligible performance drop compared to dual-view cases.

Specifically, Table II illustrates the comparison results with state-of-the-art models, where MSMV-Swin outperforms Reference [11], [19], [20], and [21] across both views and missing data scenarios. p. 4
Where the paper says it

At the same time, it is observed that the performance degradation resulting from missing views is quite negligible, illustrating the robustness of the proposed pipeline. Finally, Table II illustrates the comparison results with recent state-of-the-art [11], [19], [20] further corroborating the superiority of the proposed MSMV-Swin framework.

Conclusion

The MSMV-Swin framework successfully integrates localized and contextual characteristics of tissues surrounding tumors by leveraging SAM to extract detailed multi-scale information from both CC and MLO views. The use of two feature fusion strategies ensures robustness against missing views without relying on synthetic data. p. 4

Where the paper says it

In other words, by leveraging SAM, the framework extracts detailed multi-scale information from both CC and MLO views, effectively capturing both localized and contextual features. Additionally, two feature fusion strategies were employed to ensure robustness by compensating for missing views without relying on synthetic data.

Experimental results underscore the potential of this framework to advance reliable and comprehensive breast cancer diagnostics in real-world clinical settings. p. 4
Where the paper says it

Notably, it achieves superior accuracy and robustness even in scenarios where only single-view mammograms are available, with a negligible performance drop compared to dual-view cases. These findings underscore the potential of the MSMV-Swin framework to advance reliable and comprehensive breast cancer diagnostics in real-world clinical settings.

REFERENCES

[1] E. I. Obeagu and G. p. 4

Where the paper says it

REFERENCES [1] E. Obeagu and G.

U. Obeagu, “Breast cancer: A review of risk factors and diagnosis,” Medicine, vol. p. 4
Where the paper says it

Obeagu and G. Obeagu, “Breast cancer: A review of risk factors and diagnosis,” Medicine, vol.

103, no. 3, p. e36905, 2024

[2] A. Jouirou, A. p. 4

Where the paper says it

e36905, 2024. Jouirou, A.

Baazaoui, and W. p. 4
Where the paper says it

Jouirou, A. Baazaoui, and W.

Barhoumi, “Multi-view information fusion in mammograms: A comprehensive overview,” Information Fusion, vol. p. 4
Where the paper says it

Baazaoui, and W. Barhoumi, “Multi-view information fu- ˆ sion in mammograms: A comprehensive overview,” Information Fusion, vol.

52, pp. 308–321, 2019

[4] L. Sun, J. Wang, Z. Hu, Y. Xu, and Z. Cui, “Multi-view convolutional neural networks for mammographic image classification,” IEEE Access, vol. p. 4

Where the paper says it

165 724–165 733, 2019. Cui, “Multi-view convolutional neural networks for mammographic image classification,” IEEE Access, vol.

7, pp. 126 273–126 282, 2019

[5] X. Zhao, L. Yu, and X. Wang, “Cross-view attention network for breast cancer screening from multi-view mammograms,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). p. 4

Where the paper says it

126 273–126 282, 2019. Wang, “Cross-view attention network for breast cancer screening from multi-view mammograms,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP).

IEEE, 2020, pp. p. 4
Where the paper says it

Wang, “Cross-view attention network for breast cancer screening from multi-view mammograms,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2020, pp.

1050–1054

[6] X. Chen, K. Zhang, N. Abdoli, P. W. Gilley, X. Wang, H. Liu, B. Zheng, and Y. p. 4

Where the paper says it

1050–1054. Zheng, and Y.

Qiu, “Transformers improve breast cancer diagnosis from unregistered multi-view mammograms,” Diagnostics, vol. p. 4
Where the paper says it

Zheng, and Y. Qiu, “Transformers improve breast cancer diagnosis from unregistered multi-view mammograms,” Diagnostics, vol.

12, no 7 p 1549

[8] N. Wu, J. Phang, J. Park, Y. Shen, Z. Huang, M. Zorin et al., “Deep neural networks improve radiologists’ performance in breast cancer screening,” IEEE transactions on medical imaging, vol 39 no 4 pp 1184–1194

[9] G. Van Tulder, Y. p. 4

Where the paper says it

1184–1194, 2019. Van Tulder, Y.

Tong, and E. p. 4
Where the paper says it

Van Tulder, Y. Tong, and E.

Marchiori, “Multi-view analysis of unregistered medical images using cross-view transformers,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, France

[10] Z. Liu, Y. Lin, Y. Cao, H. Hu et al., “Swin transformer: Hierarchical vision transformer using shifted windows,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp 10 012–10 022

[11] S. Sarker, P. Sarker, G. Bebis, and A. p. 4

Where the paper says it

10 012–10 022. Bebis, and A.

Tavakkoli, “Mv-swin-t: Mammogram classification with multi-view swin transformer,” arXiv preprint arXiv:2402. p. 4
Where the paper says it

Bebis, and A. Tavakkoli, “Mv-swin-t: Mammogram classification with multi-view swin transformer,” arXiv preprint arXiv:2402.

16298, 2024

[13] A. Kirillov, E. p. 4

Where the paper says it

7948, 2020. Kirillov, E.

p. 4
Where the paper says it

7948, 2020. Kirillov, E.

Mintun, N. Ravi et al., “Segment anything,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp 4015–4026

[14] R. S. Lee, F. Gimenez, A. p. 5

Where the paper says it

Gimenez, A. Gorovoy, and D.

p. 5
Where the paper says it

Gimenez, A. Gorovoy, and D.

Hoogi et al., “A curated mammography data set for use in computer-aided detection and diagnosis research,” Scientific data, vol 4 no 1 pp 1–9, 2017

[15] H. Wang, Y. Chen, C. Ma et al., “Multimodal learning with missing modality via shared-specific feature modelling,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp 15 878–15 887

[16] Y. Chen, Y. Pan, Y. Xia, and Y. p. 5

Where the paper says it

15 878–15 887. Xia, and Y.

Yuan, “Disentangle first, then distill: A unified framework for missing modality imputation and alzheimer’s disease diagnosis,” IEEE Transactions on Medical Imaging, vol 42 no 12 pp 3566–3578

[17] F. Ma, X. Xu, S.-L. Huang, and L. p. 5

Where the paper says it

3566–3578, 2023. Huang, and L.

Zhang, “Maximum likelihood estimation for multimodal learning with missing modality,” arXiv preprint arXiv:2108. p. 5
Where the paper says it

Huang, and L. Zhang, “Maximum likelihood estimation for multimodal learning with missing modality,” arXiv preprint arXiv:2108.

10513, 2021

[18] D. Shi, L. Zhu, J. Li et al., “Incomplete crossmodal retrieval with deep correlation transfer,” ACM Transactions on Multimedia Computing, Communications and Applications, vol 20 no 5 pp 1–21, 2024

[19] K. L. dos Santos and M. p. 5

Where the paper says it

1–21, 2024. dos Santos and M.

P. dos Santos Silva, “Deep cross-training: An approach to improve deep neural network classification on mammographic images,” Expert Systems with Applications, vol 238 p 122142, 2024

[20] G. I. Quintana, Z. p. 5

Where the paper says it

122142, 2024. Quintana, Z.

Li, L. Vancamberg et al. p. 5
Where the paper says it

Quintana, Z. Vancamberg, M.

, “Exploiting patch sizes and resolutions for multi-scale deep learning in mammogram image classification,” Bioengineering, vol 10 no 5 p 534, 2023

[21] S. Yang, C. Zhang, Q. Zang et al., “Mammo-clustering: A weakly supervised multiview global-local context clustering network for detection and classification in mammography,” arXiv preprint arXiv:2409. p. 5

Where the paper says it

Liang et al. , “Mammo-clustering: A weakly supervised multiview global-local context clustering network for detection and classification in mammography,” arXiv preprint arXiv:2409.

14876, 2024

[22] H. Allaoui, Y. p. 5

Where the paper says it

14876, 2024. Allaoui, Y.

Alj, and Y. p. 5
Where the paper says it

Allaoui, Y. Alj, and Y.

Ameskine, “Hybridmammonet: A hybrid cnnvit architecture for multi-view mammography image classification,” in 2024 IEEE 12th International Symposium on Signal, Image, Video and Communications (ISIVC). p. 5
Where the paper says it

Alj, and Y. Ameskine, “Hybridmammonet: A hybrid cnnvit architecture for multi-view mammography image classification,” in 2024 IEEE 12th International Symposium on Signal, Image, Video and Communications (ISIVC).

IEEE, 2024, pp 1–6

[23] L. Liao and E. p. 5

Where the paper says it

IEEE, 2024, pp. Liao and E.

M. Aagaard, “An open codebase for enhancing transparency in deep learning-based breast cancer diagnosis utilizing cbisddsm data,” Scientific Reports, vol 14 no 1 p 27318, 2024

[7] G. Carneiro, J. p. 4

Where the paper says it

1549, 2022. Carneiro, J.

Nascimento, and A. p. 4
Where the paper says it

Carneiro, J. Nascimento, and A.

P. Bradley, “Deep learning models for classifying mammogram exams containing unregistered multi-view images and segmentation maps of lesions,” Deep learning for medical image analysis, pp 321–339, 2017

[12] P. Afshar, A. Oikonomou, F. p. 4

Where the paper says it

16298, 2024. Oikonomou, F.

Naderkhani et al. p. 4
Where the paper says it

Oikonomou, F. Naderkhani, P.

, “3d-mcn: a 3d multi-scale capsule network for lung nodule malignancy prediction,” Scientific reports, vol 10 no 1 p 7948, 2020

[13] A. Kirillov, E. Mintun, N. Ravi et al., “Segment anything,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp 4015–4026

[14] R. S. Lee, F. Gimenez, A. Hoogi et al., “A curated mammography data set for use in computer-aided detection and diagnosis research,” Scientific data, vol 4 no 1 pp 1–9, 2017

[2503.13309] The paper is titled Integrating AI for Human-Centric Breast Cancer Diagnostics: A Multi-Scale and Multi-View Swin Transformer Framework

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improve diagnostic robustness by incorporating multi-scale information that accounts for tumor-specific regions and the spatial characteristics of tissues surrounding the tumor. This allows radiologists to analyze mammograms more effectively by considering both localized tumor details and their broader anatomical context.

  2. Enhance handling of incomplete clinical data by implementing a feature zero padding strategy to provide robustness against missing views during test time. This ensures that no invalid data is propagated through the network while allowing the available view to contribute fully to the final feature representation.

  3. Improve feature extraction by leveraging the Segment Anything Model (SAM) to isolate breast lobes, which reduces background noise and enables comprehensive feature extraction. This capability allows for an extensive examination of both the localized tumor areas and the broader anatomical features, including their position, shape, and relative size within the breast anatomy.

  4. Enhance classification accuracy by employing dual feature fusion strategies: Max-Pooling Fusion or Convolution-based Fusion. These methods are designed to combine features from segmented and cropped images across CC and MLO views to capture the most critical features or leverages the robust, spatial feature processing of convolutional networks, respectively.

  5. Improve model generalization by utilizing a loss function that penalizes complex examples less severely on decision boundaries, specifically by using a binary cross-entropy with a label smoothing loss function. This helps improve model generalization by penalizing the model less severely on complex examples and smoothing the decision boundary.

Abstract

Despite advancements in Computer-Aided Diagnosis (CAD) systems, breast cancer remains one of the leading causes of cancer-related deaths among women worldwide. Recent breakthroughs in Artificial Intelligence (AI) have shown significant promise in development of advanced Deep Learning (DL) architectures for breast cancer diagnosis through mammography. In this context, the paper focuses on the integration of AI within a Human-Centric workflow to enhance breast cancer diagnostics. Key challenges are, however, largely overlooked such as reliance on detailed tumor annotations and susceptibility to missing views, particularly during test time. To address these issues, we propose a hybrid, multi-scale and multi-view Swin Transformer-based framework (MSMV-Swin) that enhances diagnostic robustness and accuracy. The proposed MSMV-Swin framework is designed to work as a decision-support tool, helping radiologists analyze multi-view mammograms more effectively. More specifically, the MSMV-Swin framework leverages the Segment Anything Model (SAM) to isolate the breast lobe, reducing background noise and enabling comprehensive feature extraction. The multi-scale nature of the proposed MSMV-Swin framework accounts for tumor-specific regions as well as the spatial characteristics of tissues surrounding the tumor, capturing both localized and contextual information. The integration of contextual and localized data ensures that MSMV-Swin's outputs align with the way radiologists interpret mammograms, fostering better human-AI interaction and trust. A hybrid fusion structure is then designed to ensure robustness against missing views, a common occurrence in clinical practice when only a single mammogram view is available.

Sources

Related papers