pith. sign in

arxiv: 2606.06537 · v1 · pith:GIAE56CNnew · submitted 2026-06-03 · 🧬 q-bio.QM · cs.CV· eess.IV

DSU-Net: An Attention-Enhanced Dense Skip U-Net for Breast Lesion Segmentation in Mammographic Images

Reza Bozorgpour , Mohammadreza Soltany Sadrabadi This is my paper

Pith reviewed 2026-06-28 02:17 UTC · model grok-4.3

classification 🧬 q-bio.QM cs.CVeess.IV
keywords breastlesionlosssegmentationaccuratedelineationdensedsu-net
0
0 comments X

The pith

DSU-Net adds dense skip connections and attention mechanisms to U-Net to segment breast lesions in mammograms, reaching a Dice score of 0.9421 on CBIS-DDSM validation data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces DSU-Net, a modified U-Net that inserts dense skip connections to preserve spatial details across layers and attention blocks to emphasize lesion boundaries during feature extraction. It trains this model on the CBIS-DDSM mammogram dataset with a combined Dice, focal, and cross-entropy loss to manage the extreme imbalance between lesion pixels and background. A sympathetic reader would care because consistent automated outlines could reduce the inter-observer differences that currently affect radiologist readings in breast cancer screening.

Core claim

The DSU-Net architecture integrates dense skip connections and attention mechanisms into the standard U-Net to improve feature propagation and lesion boundary delineation; when trained with a composite loss on the CBIS-DDSM dataset it produces a Dice Similarity Coefficient of 0.9421, an Intersection over Union of 0.8905, an accuracy of 0.9711, and an AUC-ROC of 0.9878 on the validation set, with qualitative results showing accurate outlines across lesions of different sizes and shapes.

What carries the argument

DSU-Net (attention-enhanced Dense Skip U-Net), which uses dense skip connections to maintain spatial information and attention gates to focus computation on lesion regions.

If this is right

  • The model produces consistent lesion outlines that can assist radiologists by reducing variability in boundary placement.
  • The composite loss successfully counters foreground-background imbalance during training.
  • Qualitative checks confirm the network handles lesions of varying sizes and morphologies without major boundary errors.
  • Quantitative metrics show strong separation between lesion and background pixels.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct test on images from a second scanner vendor would show whether the reported numbers depend on the specific acquisition characteristics of CBIS-DDSM.
  • The same dense-skip-plus-attention pattern could be inserted into other encoder-decoder backbones for tasks such as lung nodule segmentation in CT.
  • If attention maps are inspected after training they might reveal which image features the model uses to decide lesion extent, offering a route to clinical interpretability.
  • The absence of cross-validation across multiple random splits leaves open the possibility that the single reported validation number overstates typical performance.
  • keywords:[
  • breast lesion segmentation
  • mammography
  • U-Net

Load-bearing premise

The high performance numbers measured on the CBIS-DDSM validation split will hold for mammograms acquired on different scanners or from different patient populations.

What would settle it

Running the trained DSU-Net on an external mammogram collection from a different imaging center and obtaining a Dice score below 0.85.

read the original abstract

Breast cancer remains one of the leading causes of cancer-related mortality among women worldwide, making early detection essential for effective treatment. Mammography is the primary screening modality; however, accurate delineation of suspicious lesions remains challenging and subject to inter-observer variability. Automated segmentation methods can assist radiologists by providing consistent and efficient lesion localization. This study presents DSU-Net, an attention-enhanced Dense Skip U-Net architecture for automated breast lesion segmentation in mammographic images. The proposed framework integrates dense skip connections and attention mechanisms to improve feature propagation, preserve spatial information, and enhance lesion boundary delineation. Experiments were conducted using the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS-DDSM). To address severe foreground-background imbalance, a composite loss function combining Dice loss, focal loss, and binary cross-entropy loss was employed during training. The proposed model achieved a Dice Similarity Coefficient of 0.9421, an Intersection over Union of 0.8905, an accuracy of 0.9711, and an AUC-ROC of 0.9878 on the validation dataset. Qualitative evaluation demonstrated accurate delineation of lesions with varying sizes and morphologies, while quantitative results confirmed robust discrimination between lesion and background regions. These findings demonstrate that DSU-Net provides accurate and reliable breast lesion segmentation in mammographic images and highlights the potential of attention-guided deep learning for computer-aided breast cancer screening and diagnosis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes DSU-Net, an attention-enhanced Dense Skip U-Net for breast lesion segmentation in mammograms. It integrates dense skip connections and attention mechanisms into a U-Net backbone, trains on the CBIS-DDSM dataset using a composite loss (Dice + focal + BCE) to address class imbalance, and reports validation-set metrics of DSC 0.9421, IoU 0.8905, accuracy 0.9711, and AUC-ROC 0.9878, claiming accurate delineation across lesion sizes and morphologies.

Significance. If the numerical claims can be substantiated with proper experimental controls, the architecture offers a plausible incremental improvement for CAD segmentation tools in mammography. The composite loss and attention modules address known challenges in lesion boundary precision and foreground-background imbalance, but the absence of verifiable generalization evidence currently prevents any assessment of broader impact.

major comments (2)
  1. [Abstract] Abstract: The headline metrics (DSC 0.9421, IoU 0.8905, Acc 0.9711, AUC 0.9878) are reported on the 'validation dataset' with no description of the train-validation partitioning procedure, whether validation data informed hyper-parameter selection or early stopping, or results on any held-out test partition or external cohort. This information is load-bearing for interpreting the numbers as evidence of generalization rather than in-sample performance.
  2. [Abstract] Abstract and Methods (implied): No baseline comparisons, ablation studies on the attention or dense-skip components, or statistical significance tests against prior U-Net variants are supplied, so the contribution of the proposed modifications cannot be isolated from the reported scores.
minor comments (1)
  1. [Abstract] The abstract states that experiments were conducted on CBIS-DDSM but supplies no information on image preprocessing, patch extraction, or augmentation strategy; these details belong in the methods section for reproducibility.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive feedback on the need for greater experimental transparency and comparative analysis. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline metrics (DSC 0.9421, IoU 0.8905, Acc 0.9711, AUC 0.9878) are reported on the 'validation dataset' with no description of the train-validation partitioning procedure, whether validation data informed hyper-parameter selection or early stopping, or results on any held-out test partition or external cohort. This information is load-bearing for interpreting the numbers as evidence of generalization rather than in-sample performance.

    Authors: We agree that the partitioning procedure and validation strategy require explicit description. The revised manuscript will include a detailed account of the CBIS-DDSM train-validation split (including proportions and any stratification), confirmation that the reported validation set was held out from hyperparameter tuning and early stopping decisions, and clarification of the cross-validation or monitoring protocol used. Results on a separate held-out test partition or external cohort are not available in the current study. revision: partial

  2. Referee: [Abstract] Abstract and Methods (implied): No baseline comparisons, ablation studies on the attention or dense-skip components, or statistical significance tests against prior U-Net variants are supplied, so the contribution of the proposed modifications cannot be isolated from the reported scores.

    Authors: We acknowledge the lack of ablation studies, baseline comparisons, and statistical tests in the submitted version. The revised manuscript will incorporate ablation experiments that isolate the attention modules and dense skip connections, direct numerical comparisons against standard U-Net and relevant prior variants, and statistical significance testing (e.g., Wilcoxon or t-tests across repeated runs) to quantify the incremental benefit of the proposed components. revision: yes

standing simulated objections not resolved
  • Results on a held-out test partition or external cohort are not present in the study and cannot be supplied without new data collection.

Circularity Check

0 steps flagged

No circularity: empirical ML paper reports measured validation metrics with no derivation chain

full rationale

The paper proposes DSU-Net, an attention-enhanced U-Net variant, and states that it 'achieved a Dice Similarity Coefficient of 0.9421...' on the CBIS-DDSM validation dataset after training with a composite loss. No first-principles derivation, uniqueness theorem, ansatz, or predictive claim is advanced that could reduce to its own inputs by construction. The reported numbers are direct post-training evaluation results on a public dataset partition, which is standard empirical practice and does not match any of the enumerated circularity patterns. No self-citation load-bearing step or fitted-input-called-prediction is present in the provided text.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The central claim rests on empirical performance after training a high-capacity neural network; the model contains thousands of learned weights plus unspecified loss-component weights and training choices that are not derived from first principles.

free parameters (2)
  • loss component weights
    The composite loss (Dice + focal + BCE) requires weighting coefficients whose values are not stated in the abstract.
  • network hyperparameters
    Learning rate, batch size, attention module dimensions, and number of dense blocks are free choices that affect the final metrics.
axioms (1)
  • domain assumption The CBIS-DDSM validation images are representative of the distribution on which the model will be deployed.
    The claim of reliable segmentation depends on this untested generalization assumption.

pith-pipeline@v0.9.1-grok · 5805 in / 1376 out tokens · 34929 ms · 2026-06-28T02:17:44.665588+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references

  1. [1]

    Introduction Breast cancer remains one of the leading health concerns affecting women and is the most frequently diagnosed cancer among females in the United States [1, 2]. Improving patient outcomes largely depends on the early detection of malignant breast lesions, as diagnosis at an earlier stage can significantly reduce disease progression and mortali...

  2. [2]

    Motivation Breast lesion segmentation in mammographic images remains challenging due to large variations in lesion morphology, low contrast between lesions and surrounding tissue, and the severe class imbalance between lesion and background regions. Although U -Net and its variants have achieved promising results, many existing approaches struggle to effe...

  3. [3]

    Materials and Methods 3.1. Dataset The proposed segmentation framework was developed using the Curated Breast Imaging Subset of the Digital Database for Screening Mammography (CBIS -DDSM), a publicly available and extensively used benchmark dataset for breast cancer research. The dataset con tains digitized mammographic images together with expert annotat...

  4. [4]

    The training configuration and hyperparameters used in this study are summarized in Table 4

    Results The proposed DSU -Net model demonstrated stable convergence during training on the CBIS -DDSM dataset. The training configuration and hyperparameters used in this study are summarized in Table 4. Although training was configured for a maximum of 100 epochs, early stopping terminated the process after 24 epochs. The best validation performance, pre...

  5. [5]

    In the CBIS-DDSM dataset, lesion pixels typically occupy only a small portion of the image while the majority of pixels belong to the background

    Discussion The results demonstrate that the proposed DSU-Net architecture is capable of accurately segmenting breast lesions from mammographic regions of interest despite the substantial class imbalance inherent to mammography datasets. In the CBIS-DDSM dataset, lesion pixels typically occupy only a small portion of the image while the majority of pixels ...

  6. [6]

    Limitation Several limitations of the present study should be acknowledged. First, model development and evaluation were performed using the CBIS-DDSM dataset, which, although widely used for breast imaging research, may not fully represent the diversity of mammographic images encountered in clinical practice. Variations in imaging protocols, acquisition ...

  7. [7]

    CA: a cancer journal for clinicians, 2022

    Siegel, R.L., et al., Cancer statistics, 2022. CA: a cancer journal for clinicians, 2022. 72(1): p. 7 – 33

    2022

  8. [8]

    IEEE journal of biomedical and health informatics, 2017

    Yap, M.H., et al., Automated breast ultrasound lesions detection using convolutional neural networks. IEEE journal of biomedical and health informatics, 2017. 22(4): p. 1218–1226

    2017

  9. [9]

    Computer Methods and Programs in Biomedicine, 2023

    Gerbasi, A., et al., DeepMiCa: Automatic segmentation and classification of breast MIcroCAlcifications from mammograms. Computer Methods and Programs in Biomedicine, 2023. 235: p. 107483

    2023

  10. [10]

    Radiography, 2022

    Aristokli, N., et al., Comparison of the diagnostic performance of Magnetic Resonance Imaging (MRI), ultrasound and mammography for detection of breast cancer based on tumor type, breast density and patient's history: A review. Radiography, 2022. 28(3): p. 848–856

    2022

  11. [11]

    Clinical Breast Cancer, 2023

    Kim, Y .E., et al., The accuracy of mammography, ultrasound, and magnetic resonance imaging for the measurement of invasive breast cancer with extensive intraductal components. Clinical Breast Cancer, 2023. 23(1): p. 45–53

    2023

  12. [12]

    Dai, and J

    Chen, G., Y . Dai, and J. Zhang, C-Net: Cascaded convolutional neural network with global guidance and refinement residuals for breast ultrasound images segmentation. Computer Methods and Programs in Biomedicine, 2022. 225: p. 107086

    2022

  13. [13]

    IEEE Journal of Biomedical and Health Informatics, 2023

    Hu, K., et al., Boundary-guided and region-aware network with global scale-adaptive for accurate segmentation of breast tumors in ultrasound images. IEEE Journal of Biomedical and Health Informatics, 2023. 27(9): p. 4421–4432

    2023

  14. [14]

    Computers in biology and medicine, 2021

    Webb, J.M., et al., Comparing deep learning-based automatic segmentation of breast masses to expert interobserver variability in ultrasound imaging. Computers in biology and medicine, 2021. 139: p. 104966

    2021

  15. [15]

    Computer methods and programs in biomedicine,

    Cheng, Z., et al., DSGMFFN: Deepest semantically guided multi-scale feature fusion network for automated lesion segmentation in ABUS images. Computer methods and programs in biomedicine,

  16. [16]

    Cohen, and Y .C

    Van Sloun, R.J., R. Cohen, and Y .C. Eldar, Deep learning in ultrasound imaging. Proceedings of the IEEE, 2019. 108(1): p. 11–29

    2019

  17. [17]

    Computer Methods and Programs in Biomedicine, 2023

    Zhong, S., et al., MsGoF: Breast lesion classification on ultrasound images by multi -scale gradational-order fusion framework. Computer Methods and Programs in Biomedicine, 2023. 230: p. 107346

    2023

  18. [18]

    BAGNet: bidirectional aware guidance network for malignant breast lesions segmentation

    Chen, G., et al. BAGNet: bidirectional aware guidance network for malignant breast lesions segmentation. in 2022 7th Asia-Pacific Conference on Intelligent Robot Systems (ACIRS) . 2022. IEEE

    2022

  19. [19]

    Computers in biology and medicine, 2022

    Khaled, R.a., et al., A U-Net Ensemble for breast lesion segmentation in DCE MRI. Computers in biology and medicine, 2022. 140: p. 105093

    2022

  20. [20]

    Breast-FRUNET: Automated Breast Cancer Classification Using X-Ray Images with Faster R -CNN and U -Net

    Aashish, G., et al. Breast-FRUNET: Automated Breast Cancer Classification Using X-Ray Images with Faster R -CNN and U -Net. in 2025 International Conference on Computing Technologies (ICOCT). 2025. IEEE

    2025

  21. [21]

    Hiwa, H. and A.A. Abdulla, Early Breast Cancer Detection for Mammogram Images using U-Net Deep Learning Model. Zanin Journal of Science and Engineering, 2025. 1(4): p. 144–162

    2025

  22. [22]

    Kalpana, V . and S. Thilagamani, TOMPEI-CMMD: An Adaptive Radiological Feature Analysis Using Unitary U-Net for High-Precision Mammography Segmentation. Neurocomputing, 2025: p. 131982

    2025

  23. [23]

    A comparative analysis of breast cancer detection and diagnosis using data visualization and machine learning applications

    Ak, M.F. A comparative analysis of breast cancer detection and diagnosis using data visualization and machine learning applications. in Healthcare. 2020. MDPI

    2020

  24. [24]

    Chaumrattanakul, and S.S

    Ilesanmi, A.E., U. Chaumrattanakul, and S.S. Makhanov, Methods for the segmentation and classification of breast ultrasound images: a review. Journal of ultrasound, 2021. 24(4): p. 367 – 382

    2021

  25. [25]

    Computers in biology and medicine, 2023

    Ma, Z., et al., ATFE-Net: axial transformer and feature enhancement -based CNN for ultrasound breast mass segmentation. Computers in biology and medicine, 2023. 153: p. 106533

    2023

  26. [26]

    Medical image analysis, 2022

    Huang, R., et al., Boundary-rendering network for breast lesion segmentation in ultrasound images. Medical image analysis, 2022. 80: p. 102478

    2022

  27. [27]

    Biomedical Signal Processing and Control, 2023

    Lyu, Y ., et al., AMS-P AN: Breast ultrasound image segmentation model combining attention mechanism and multi -scale features. Biomedical Signal Processing and Control, 2023. 81: p. 104425

    2023

  28. [28]

    Artificial Intelligence Review, 2023

    Pengiran Mohamad, D.N.F., et al., Transition of traditional method to deep learning based computer-aided system for breast cancer using automated breast ultrasound system (abus) images: a review. Artificial Intelligence Review, 2023. 56(12): p. 15271–15300

    2023

  29. [29]

    Computers in Biology and Medicine, 2023

    Ru, J., et al., Attention guided neural ODE network for breast tumor segmentation in medical images. Computers in Biology and Medicine, 2023. 159: p. 106884

    2023

  30. [30]

    Aruldharani, and M

    Dharani, S., S. Aruldharani, and M. Dharani Priya. Multi-Modal Swin-Transformer Framework for Preliminary Breast Cancer Screening Through Architectural Distortion Analysis . in 2026 5th International Conference on Sentiment Analysis and Deep Learning (ICSADL). 2026. IEEE

    2026

  31. [31]

    Electronic Imaging, 2026

    Guo, Q., et al., Local Attention and Detail -enhanced Network for Mass Segmentation in Whole Mammograms. Electronic Imaging, 2026. 38: p. 1–7

    2026

  32. [32]

    Polaka, and K

    Rajasekaran, M., A.R. Polaka, and K. Reddy. AI-Assisted Breast Cancer Diagnosis Using Vision Transformer. in 2026 IEEE International Conference on AI Engineering and Innovations (AIEI)

    2026

  33. [33]

    Acharya, and F

    Meiburger, K.M., U.R. Acharya, and F. Molinari, Automated localization and segmentation techniques for B-mode ultrasound images: A review. Computers in biology and medicine, 2018. 92: p. 210–235

    2018

  34. [34]

    Fischer, and T

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

    2015

  35. [35]

    Biomedical signal processing and control, 2020

    Byra, M., et al., Breast mass segmentation in ultrasound with selective kernel U-Net convolutional neural network. Biomedical signal processing and control, 2020. 61: p. 102027

    2020

  36. [36]

    IEEE Transactions on Medical Imaging, 2022

    Chen, G., et al., AAU-net: an adaptive attention U-net for breast lesions segmentation in ultrasound images. IEEE Transactions on Medical Imaging, 2022. 42(5): p. 1289–1300

    2022

  37. [37]

    Plos one, 2021

    Guo, Y ., et al., Segmentation and recognition of breast ultrasound images based on an expanded U-Net. Plos one, 2021. 16(6): p. e0253202

    2021

  38. [38]

    PloS one, 2018

    Kumar, V ., et al., Automated and real -time segmentation of suspicious breast masses using convolutional neural network. PloS one, 2018. 13(5): p. e0195816

    2018

  39. [39]

    SN Computer Science, 2025

    Abdel-Wahab, A., et al., Deep learning framework for breast cancer detection and segmentation using EfficientNet and U-Net with hyperparameter optimization. SN Computer Science, 2025. 6(5): p. 528

    2025

  40. [40]

    Complex & Intelligent Systems, 2024

    El-Banby, G.M., et al., Automated abnormalities detection in mammography using deep learning. Complex & Intelligent Systems, 2024. 10(5): p. 7279–7295

    2024

  41. [41]

    Segmentation in mammographic images based on the U -NET model

    Fazilov, S., et al. Segmentation in mammographic images based on the U -NET model. in AIP Conference Proceedings. 2024. AIP Publishing LLC

    2024

  42. [42]

    Punn, N.S. and S. Agarwal, RCA-IUnet: a residual cross-spatial attention-guided inception U-Net model for tumor segmentation in breast ultrasound imaging. Machine Vision and Applications,

  43. [43]

    Biomedical Signal Processing and Control, 2022

    Yan, Y ., et al., Accurate segmentation of breast tumors using AE U -net with HDC model in ultrasound images. Biomedical Signal Processing and Control, 2022. 72: p. 103299

    2022

  44. [44]

    Biomedical Signal Processing and Control, 2021

    Tong, Y ., et al., Improved U-net MALF model for lesion segmentation in breast ultrasound images. Biomedical Signal Processing and Control, 2021. 68: p. 102721

    2021

  45. [45]

    IEEE Access, 2021

    Li, J., et al., Multi-scale fusion U-net for the segmentation of breast lesions. IEEE Access, 2021. 9: p. 137125–137139

    2021

  46. [46]

    IEEE transactions on medical imaging, 2021

    Ning, Z., et al., SMU-Net: Saliency -guided morphology -aware U -Net for breast lesion segmentation in ultrasound image. IEEE transactions on medical imaging, 2021. 41(2): p. 476 – 490

    2021