HADS-Net:A Hybrid Attention-Augmented Dual-Stream Network with Physics-Informed Augmentation for Breast Ultrasound Image Classification

Blessing Nwamaka Iduh; Chinedu Emmanuel Mbonu; Doris Chinedu Asogwa; Joseph Ikechukwu Odo

arxiv: 2605.20536 · v1 · pith:A6I3QXL5new · submitted 2026-05-19 · 💻 cs.CV

HADS-Net:A Hybrid Attention-Augmented Dual-Stream Network with Physics-Informed Augmentation for Breast Ultrasound Image Classification

Chinedu Emmanuel Mbonu , Blessing Nwamaka Iduh , Joseph Ikechukwu Odo , Doris Chinedu Asogwa This is my paper

Pith reviewed 2026-05-21 06:42 UTC · model grok-4.3

classification 💻 cs.CV

keywords breast ultrasound classificationdual-stream networkphysics-informed augmentationcross-attention fusionlesion boundary detectionEfficientNetBUSI dataset

0 comments

The pith

A dual-stream network pairs physics-simulated ultrasound artifacts with explicit boundary edges and cross-attention to classify breast ultrasound images at 96.58 percent accuracy.

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

The paper presents HADS-Net to address the problem of classifying breast ultrasound scans into benign, malignant, and normal categories despite speckle noise, shadowing, and visual overlap between classes. It runs two parallel pathways: one applies simulated clinical ultrasound distortions to a pretrained EfficientNet model to capture texture, while the other computes Sobel edge maps to isolate lesion boundaries that clinicians rely on most. These representations are combined through a cross-attention module so texture features can draw selectively on boundary information, and the fused vector is passed to an MLP trained with class-weighted focal loss. On the BUSI dataset the model records 96.58 percent accuracy, 0.9978 macro ROC-AUC, and 0.9654 macro F1, with no malignant cases labeled normal. The results indicate that embedding ultrasound acquisition physics and boundary emphasis into the architecture improves separation of the three classes over generic single-stream approaches.

Core claim

HADS-Net processes each breast ultrasound image through a texture stream that receives physics-informed augmentations simulating speckle noise, acoustic shadowing, and gain variation before feature extraction with pretrained EfficientNet-B3, and a parallel boundary stream that feeds Sobel edge maps into a lightweight CNN; both streams project to 512 dimensions, a cross-attention module lets the texture pathway query boundary cues, and the joint representation is classified by an MLP under adaptive class-weighted focal loss, yielding 96.58 percent accuracy, 0.9978 macro ROC-AUC, 0.9654 macro F1, and per-class F1 scores of 0.970, 0.951, and 0.976 on the BUSI dataset with zero malignant-to-weak

What carries the argument

The cross-attention fusion module that lets the global texture stream selectively query and integrate local boundary features extracted from Sobel edge maps.

If this is right

Modality-specific artifact simulation combined with explicit boundary extraction produces reliable separation of malignant from normal cases.
Cross-attention between texture and edge streams integrates complementary cues that single-stream models miss.
Class-weighted focal loss with the dual representation reduces the impact of visual ambiguity between benign and malignant lesions.
The architecture can be retrained on other ultrasound tasks once the same physics simulation parameters are calibrated to the new acquisition setting.

Where Pith is reading between the lines

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

If the physics augmentations generalize, the same dual-stream pattern could improve classification in other ultrasound applications such as liver or thyroid imaging where shadowing and speckle are also dominant.
Running the boundary stream on raw images rather than Sobel maps might reveal whether edge detection itself is the critical step or whether the attention mechanism can learn boundaries from texture alone.
Collecting a small external test set from a different ultrasound machine would directly test whether the simulated artifacts transfer without retraining.

Load-bearing premise

The simulated speckle noise, acoustic shadowing, and gain variation applied in the texture stream match the actual distribution of artifacts encountered in real clinical breast ultrasound acquisitions without introducing systematic bias.

What would settle it

Replace the physics-informed augmentations with standard random flips, rotations, and color jitter while keeping all other components fixed, then measure whether accuracy on the held-out BUSI test set falls below 95 percent or malignant-to-normal misclassifications appear.

Figures

Figures reproduced from arXiv: 2605.20536 by Blessing Nwamaka Iduh, Chinedu Emmanuel Mbonu, Doris Chinedu Asogwa, Joseph Ikechukwu Odo.

**Figure 1.** Figure 1: HADS-Net training pipeline. Input splits into Stream 1 (physics aug. + EfficientNet-B3) and Stream 2 (Sobel edges + lightweight CNN). Both streams [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

**Figure 3.** Figure 3: Train (dashed) vs. validation (solid) loss and accuracy across all five [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 2.** Figure 2: Sample BUSI images. Top: benign (well-defined oval masses). Middle: [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 4.** Figure 4: Test set evaluation. Left: best validation loss per fold (Fold 1 highlighted as global best). Centre: confusion matrix on the held-out test set. Right: [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

read the original abstract

Accurate classification of breast ultrasound images into benign, malignant, and normal categories is a critical clinical task complicated by speckle noise, acoustic shadowing, and inter-class visual ambiguity. Existing deep learning methods rely on single-stream architectures with generic augmentation that ignores ultrasound acquisition physics, and no prior method dedicates a stream to the lesion boundary features identified as the most diagnostically significant visual cue. We propose HADS-Net, a Hybrid Attention-Augmented Dual-Stream Network exploiting global texture and local boundary cues through two parallel pathways. Stream 1 applies physics-informed augmentation simulating speckle noise, acoustic shadowing, and gain variation before extracting features via pretrained EfficientNet-B3 projected to 512 dimensions. Stream 2 extracts Sobel edge maps processed by a lightweight CNN projected to the same 512-dimensional space. A cross-attention fusion module allows the texture stream to selectively query boundary features, producing a jointly optimised representation classified by an MLP trained with adaptive class-weighted focal loss. Five-fold stratified cross-validation with cosine annealing over 50 epochs is used, with the globally best checkpoint selected by lowest validation loss evaluated on a held-out test set. On the BUSI dataset, HADS-Net achieves 96.58% accuracy, macro ROC-AUC of 0.9978, macro F1 of 0.9654, and per-class F1-scores of 0.970, 0.951, and 0.976 for benign, malignant, and normal. No malignant lesion is misclassified as normal. These results confirm that modality-specific augmentation with cross-modal attention fusion is an effective strategy for ultrasound-based breast cancer 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.

Desk Editor's Note private letter to a colleague

HADS-Net adds a boundary stream and physics-based augmentations to a dual-stream setup for ultrasound classification and posts strong BUSI numbers, but the lack of baselines and augmentation checks leaves the gains hard to verify.

read the letter

The main things to know are that the authors split the model into a texture path that gets ultrasound-specific augmentations before EfficientNet-B3 and a separate Sobel-edge path, then fuse them with cross-attention before an MLP classifier. On the BUSI dataset they report 96.58% accuracy, 0.9978 macro AUC, and no malignant-to-normal errors under 5-fold CV plus held-out testing. That combination of domain-aware augmentation and explicit boundary attention is the concrete step they take beyond generic single-stream models. The architecture description is clear enough that someone could reimplement the streams and fusion module without much guesswork, and the training choices like cosine annealing and class-weighted focal loss are standard but sensible for the imbalanced medical setting. The per-class F1 scores and the zero critical errors are the results that would interest a reader working on diagnostic tools. The soft spots are the missing controls. The abstract and method summary give no baseline numbers against plain EfficientNet or prior ultrasound classifiers, no ablation tables showing what the boundary stream or the specific augmentations actually contribute, and no statistical tests or error bars. Without those it is difficult to separate the proposed design from careful hyperparameter work or a strong backbone. The physics-informed augmentations for speckle, shadowing, and gain are described but not checked against real BUSI artifact statistics, so the stress-test concern about possible distribution shift is reasonable on the evidence given. This paper is for people already working on medical ultrasound classification who want a practical recipe that injects acquisition physics and boundary cues. A reader could pull the dual-stream idea and try it on their own data, but they would still need to add their own comparisons. I would send it for peer review. The method is well-motivated, the dataset is public, and the reported performance is high enough that referees can usefully ask for the missing baselines and validation steps rather than reject outright.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes HADS-Net, a hybrid attention-augmented dual-stream network for classifying breast ultrasound images into benign, malignant, and normal categories on the BUSI dataset. Stream 1 applies physics-informed augmentations (speckle noise, acoustic shadowing, gain variation) before EfficientNet-B3 feature extraction projected to 512 dimensions; Stream 2 processes Sobel edge maps via a lightweight CNN also projected to 512 dimensions. A cross-attention fusion module integrates the streams, followed by an MLP classifier trained with adaptive class-weighted focal loss. Using 5-fold stratified cross-validation with cosine annealing over 50 epochs and a held-out test set, the method reports 96.58% accuracy, 0.9978 macro ROC-AUC, 0.9654 macro F1, per-class F1 scores of 0.970/0.951/0.976, and zero malignant-to-normal misclassifications.

Significance. If the reported performance holds under proper validation, the dual-stream design combining physics-informed texture augmentation with boundary-focused edge processing and cross-attention fusion offers a targeted approach to handling ultrasound-specific artifacts and visual ambiguity. The zero critical misclassifications and high ROC-AUC are notable strengths for potential clinical utility in breast cancer diagnosis.

major comments (3)

[Method (Stream 1 physics-informed augmentation)] Method section describing Stream 1: The physics-informed augmentations simulating speckle noise, acoustic shadowing, and gain variation are applied prior to EfficientNet-B3 without any reported calibration (e.g., parameter estimation from real BUSI images) or statistical validation (e.g., Kolmogorov-Smirnov tests on noise histograms or frequency-domain matching of shadowing patterns). This is load-bearing for the central claim, as the headline metrics and assertion that 'modality-specific augmentation with cross-modal attention fusion is effective' depend on these simulations faithfully reproducing real clinical artifact distributions.
[Results] Results section: The quantitative results (96.58% accuracy, macro ROC-AUC 0.9978, per-class F1 scores) are presented without any baseline comparisons to prior single-stream or generic-augmentation methods on the identical BUSI dataset and splits, nor ablation studies isolating the contribution of the physics augmentations, Sobel stream, or cross-attention module. This prevents assessment of whether the gains are attributable to the proposed architecture.
[Evaluation protocol] Evaluation protocol: No error bars, statistical significance tests (e.g., McNemar or paired t-tests across folds), or details on held-out test set size/composition are provided, which is necessary to support the stability of the reported macro F1 of 0.9654 and the claim of no malignant-to-normal errors under 5-fold CV.

minor comments (2)

[Abstract] The abstract states that 'no prior method dedicates a stream to the lesion boundary features' but provides no citations to the reviewed literature supporting this positioning.
[Method (feature projection)] The projection layers reducing features to 512 dimensions are mentioned for both streams but lack explicit layer dimensions, activation functions, or initialization details needed for full reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments, which have helped us identify areas to strengthen the manuscript. We provide point-by-point responses to each major comment below and commit to revisions that enhance methodological transparency, comparative analysis, and statistical rigor without altering the core contributions.

read point-by-point responses

Referee: Method section describing Stream 1: The physics-informed augmentations simulating speckle noise, acoustic shadowing, and gain variation are applied prior to EfficientNet-B3 without any reported calibration (e.g., parameter estimation from real BUSI images) or statistical validation (e.g., Kolmogorov-Smirnov tests on noise histograms or frequency-domain matching of shadowing patterns). This is load-bearing for the central claim, as the headline metrics and assertion that 'modality-specific augmentation with cross-modal attention fusion is effective' depend on these simulations faithfully reproducing real clinical artifact distributions.

Authors: We acknowledge that the manuscript does not currently include explicit calibration details or statistical validation for the augmentation parameters. The parameters were selected to reflect typical ultrasound acquisition physics as described in the literature on speckle noise modeling, shadowing effects, and gain adjustments. In the revised version, we will expand the Method section to specify the exact parameter ranges and distributions used for each augmentation type, provide justification drawn from established ultrasound imaging studies, and report statistical comparisons (including Kolmogorov-Smirnov tests on intensity histograms and frequency-domain analysis of shadowing patterns) between the augmented images and the original BUSI data to demonstrate fidelity to real clinical distributions. revision: yes
Referee: Results section: The quantitative results (96.58% accuracy, macro ROC-AUC 0.9978, per-class F1 scores) are presented without any baseline comparisons to prior single-stream or generic-augmentation methods on the identical BUSI dataset and splits, nor ablation studies isolating the contribution of the physics augmentations, Sobel stream, or cross-attention module. This prevents assessment of whether the gains are attributable to the proposed architecture.

Authors: We agree that direct baseline comparisons and ablation studies are necessary to isolate the contributions of the proposed components. The current manuscript focuses on presenting the full HADS-Net results but does not include these analyses. We will add a dedicated subsection to the Results section that reports performance of relevant baselines (including single-stream EfficientNet-B3 with standard augmentations and prior BUSI methods) using identical dataset splits and 5-fold CV protocol. We will also include comprehensive ablation experiments that remove the physics-informed augmentations, the Sobel boundary stream, and the cross-attention fusion module individually, quantifying the resulting changes in accuracy, ROC-AUC, and F1 scores to demonstrate each element's impact. revision: yes
Referee: Evaluation protocol: No error bars, statistical significance tests (e.g., McNemar or paired t-tests across folds), or details on held-out test set size/composition are provided, which is necessary to support the stability of the reported macro F1 of 0.9654 and the claim of no malignant-to-normal errors under 5-fold CV.

Authors: We recognize that reporting variability and conducting statistical tests would better substantiate the stability of the results. The manuscript currently reports aggregate metrics from 5-fold stratified cross-validation and a held-out test set but omits error bars and formal significance testing. In the revised manuscript, we will add standard deviation error bars for all reported metrics across the five folds. We will perform and report paired statistical tests, including McNemar's test for misclassification patterns and paired t-tests for metric differences across folds. We will also explicitly detail the held-out test set size, class distribution, and selection process to confirm its independence and representativeness. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical training and held-out evaluation on BUSI dataset

full rationale

The paper describes a dual-stream CNN architecture (EfficientNet-B3 texture stream with physics-informed augmentations plus Sobel-edge boundary stream) fused by cross-attention and trained with focal loss. All reported metrics (96.58% accuracy, 0.9978 ROC-AUC, per-class F1 scores) are obtained via 5-fold stratified cross-validation plus held-out test set selection by validation loss. No equations, derivations, or self-citations are invoked to obtain these quantities; the results are produced by standard supervised optimization on external data and do not reduce to any fitted parameter or prior result by construction. The physics-informed augmentations are an input preprocessing choice whose fidelity is an empirical assumption, not a definitional loop.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 0 invented entities

The performance claim depends on several modeling choices and dataset assumptions that are not independently verified beyond the reported experiments on a single public benchmark.

free parameters (3)

feature projection dimension = 512
Both streams are projected to the same 512-dimensional space before cross-attention fusion.
training epochs and schedule = 50
Cosine annealing schedule run for 50 epochs with checkpoint selection by lowest validation loss.
loss weighting parameters
Adaptive class-weighted focal loss tuned to handle class imbalance in the three-class problem.

axioms (2)

domain assumption The BUSI dataset distribution is representative of clinical breast ultrasound images for the purpose of evaluating classification performance.
All reported metrics are obtained solely on this dataset with no external or multi-center validation mentioned.
domain assumption The simulated speckle noise, acoustic shadowing and gain variation in Stream 1 sufficiently approximate real ultrasound acquisition physics.
The physics-informed augmentation is presented as the key differentiator without quantitative validation against real acquisition variability.

pith-pipeline@v0.9.0 · 5857 in / 1823 out tokens · 67661 ms · 2026-05-21T06:42:22.301792+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Stream 1 applies physics-informed augmentation simulating speckle noise, acoustic shadowing, and gain variation before feeding images through a pretrained EfficientNet-B3 backbone
IndisputableMonolith/Foundation/AlexanderDuality.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

A novel cross-attention fusion module then allows the texture stream to selectively query boundary features

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 1 internal anchor

[1]

Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,

F. Bray, J. Ferlay, I. Soerjomataram, R. L. Siegel, L. A. Torre, and A. Jemal, “Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,”CA: A Cancer Journal for Clinicians, vol. 68, no. 6, pp. 394–424, 2018. Fig. 4. Test set evaluation. Left: best validation loss per fold (Fold 1 highlighted ...

work page 2018
[2]

Dataset of breast ultrasound images,

W. Al-Dhabyani, M. Gomaa, H. Khaled, and A. Fahmy, “Dataset of breast ultrasound images,”Data in Brief, vol. 28, p. 104863, 2020

work page 2020
[3]

Enhancing breast cancer diagnosis with vision transformer-based ultrasound image classification,

A. Ashraf, A. E. Nagib, and H. Mohamed, “Enhancing breast cancer diagnosis with vision transformer-based ultrasound image classification,” inProc. 5th Novel Intelligent and Leading Emerg- ing Sciences Conference (NILES), IEEE, 2023, pp. 161–165. doi: 10.1109/NILES59815.2023.10296582

work page doi:10.1109/niles59815.2023.10296582 2023
[4]

Explaining a deep learning based breast ultrasound image classifier with saliency maps,

M. Byra, K. Dobruch-Sobczak, H. Piotrzkowska-Wroblewska, Z. Klimonda, and J. Litniewski, “Explaining a deep learning based breast ultrasound image classifier with saliency maps,” Journal of Ultrasonography, vol. 22, pp. e70–e75, 2022. doi: 10.15557/JoU.2022.0013

work page doi:10.15557/jou.2022.0013 2022
[5]

Breast UltraSound Image classification using fuzzy-rank-based ensemble network,

S. D. Deb and R. K. Jha, “Breast UltraSound Image classification using fuzzy-rank-based ensemble network,”Biomedical Signal Processing and Control, vol. 85, p. 104871, 2023

work page 2023
[6]

Vision transformers for classification of breast ultrasound images,

B. Gheflati and H. Rivaz, “Vision transformers for classification of breast ultrasound images,” inProc. 44th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 480– 483, 2022

work page 2022
[7]

Deep learning approaches for classification of breast cancer in ultrasound (US) images,

I. Pacal, “Deep learning approaches for classification of breast cancer in ultrasound (US) images,”Journal of the Institute of Science and Tech- nology, vol. 12, no. 4, pp. 1917–1927, 2022. doi: 10.21597/jist.1183679

work page doi:10.21597/jist.1183679 1917
[8]

Breast cancer classification based on convolutional neural network and image fusion approaches using ultrasound images,

M. Alotaibi, A. Aljouie, N. Alluhaidan, W. Qureshi, H. Almatar, R. Al- duhayan, B. Alsomaie, and A. Almazroa, “Breast cancer classification based on convolutional neural network and image fusion approaches using ultrasound images,”Heliyon, vol. 9, no. 11, p. e22406, 2023. doi: 10.1016/j.heliyon.2023.e22406

work page doi:10.1016/j.heliyon.2023.e22406 2023
[9]

Enhancing breast cancer segmentation and classification: An ensemble deep convolu- tional neural network and U-net approach on ultrasound images,

M. R. Islam, M. M. Rahman, M. S. Ali, A. A. N. Nafi, M. S. Alam, T. K. Godder, M. S. Miah, and M. K. Islam, “Enhancing breast cancer segmentation and classification: An ensemble deep convolu- tional neural network and U-net approach on ultrasound images,” Machine Learning with Applications, vol. 16, p. 100555, 2024. doi: 10.1016/j.mlwa.2024.100555

work page doi:10.1016/j.mlwa.2024.100555 2024
[10]

An EfficientNet integrated ResNet deep network and explainable AI for breast lesion classification from ultrasound images,

K. Jabeen, M. A. Khan, A. Hamza, H. M. Albarakati, S. Alsenan, U. Tariq, and I. Ofori, “An EfficientNet integrated ResNet deep network and explainable AI for breast lesion classification from ultrasound images,”CAAI Transactions on Intelligence Technology, vol. 10, no. 3, pp. 842–857, 2025. doi: 10.1049/cit2.12385

work page doi:10.1049/cit2.12385 2025
[11]

Improving breast cancer diagnosis in ultrasound images using deep learning with feature fusion and attention mechanism,

S. Asif, Y . Yan, B. Feng, M. Wang, Y . Zheng, T. Jiang, R. Fu, J. Yao, L. Lv, M. Song, L. Sui, Z. Yin, V . Y . Wang, and D. Xu, “Improving breast cancer diagnosis in ultrasound images using deep learning with feature fusion and attention mechanism,”Academic Radiology, vol. 32, pp. 4997–5009, 2025. doi: 10.1016/j.acra.2025.05.007

work page doi:10.1016/j.acra.2025.05.007 2025
[12]

Breast can- cer classification from ultrasound images using VGG16 model based transfer learning,

A. B. M. A. Hossain, J. K. Nisha, and F. Johora, “Breast can- cer classification from ultrasound images using VGG16 model based transfer learning,”International Journal of Image, Graphics and Sig- nal Processing (IJIGSP), vol. 15, no. 1, pp. 12–22, 2023. doi: 10.5815/ijigsp.2023.01.02

work page doi:10.5815/ijigsp.2023.01.02 2023
[13]

Advancing breast ultrasound diagnostics through hybrid deep learning models,

A. Kiran, J. V . N. Ramesh, I. S. Rahat, M. A. U. Khan, A. Hossain, and R. Uddin, “Advancing breast ultrasound diagnostics through hybrid deep learning models,”Computers in Biology and Medicine, vol. 180, p. 108962, 2024. doi: 10.1016/j.compbiomed.2024.108962

work page doi:10.1016/j.compbiomed.2024.108962 2024
[14]

An explorative analysis of SVM classifier and ResNet50 architecture on African food classification,

C. Mbonu, K. Anigbogu, D. Asogwa, and T. Belonwu, “An explorative analysis of SVM classifier and ResNet50 architecture on African food classification,”arXiv preprint arXiv:2505.13923, 2025

work page arXiv 2025
[15]

EfficientNet: Rethinking model scaling for convolutional neural networks,

M. Tan and Q. V . Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” inProc. 36th International Conference on Machine Learning (ICML), vol. 97, pp. 6105–6114, 2019

work page 2019
[16]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, pp. 5998– 6008, 2017

work page 2017
[17]

Focal loss for dense object detection,

T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” inProc. IEEE International Conference on Computer Vision (ICCV), pp. 2980–2988, 2017

work page 2017
[18]

Decoupled Weight Decay Regularization

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inarXiv preprint arXiv:1711.05101, 2017

work page internal anchor Pith review Pith/arXiv arXiv 2017
[19]

A review on evaluation metrics for data classification evaluations,

M. Hossin and M. N. Sulaiman, “A review on evaluation metrics for data classification evaluations,”International Journal of Data Mining and Knowledge Management Process, vol. 5, no. 2, pp. 1–11, 2015

work page 2015
[20]

Metrics for multi-class classifi- cation: an overview,

M. Grandini, E. Bagli, and G. Visani, “Metrics for multi-class classifi- cation: an overview,”arXiv preprint arXiv:2008.05756, 2020

work page arXiv 2008
[21]

An introduction to ROC analysis,

T. Fawcett, “An introduction to ROC analysis,”Pattern Recognition Letters, vol. 27, no. 8, pp. 861–874, 2006

work page 2006
[22]

A multi-algorithmic approach to stroke risk prediction using machine learning,

E. G. Onyedinma, D. C. Asogwa, T. S. Belonwu, and C. E. Mbonu, “A multi-algorithmic approach to stroke risk prediction using machine learning,”Journal of Engineering Research and Reports, vol. 27, no. 7, pp. 247–259, 2025. doi: 10.9734/jerr/2025/v27i71573

work page doi:10.9734/jerr/2025/v27i71573 2025

[1] [1]

Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,

F. Bray, J. Ferlay, I. Soerjomataram, R. L. Siegel, L. A. Torre, and A. Jemal, “Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,”CA: A Cancer Journal for Clinicians, vol. 68, no. 6, pp. 394–424, 2018. Fig. 4. Test set evaluation. Left: best validation loss per fold (Fold 1 highlighted ...

work page 2018

[2] [2]

Dataset of breast ultrasound images,

W. Al-Dhabyani, M. Gomaa, H. Khaled, and A. Fahmy, “Dataset of breast ultrasound images,”Data in Brief, vol. 28, p. 104863, 2020

work page 2020

[3] [3]

Enhancing breast cancer diagnosis with vision transformer-based ultrasound image classification,

A. Ashraf, A. E. Nagib, and H. Mohamed, “Enhancing breast cancer diagnosis with vision transformer-based ultrasound image classification,” inProc. 5th Novel Intelligent and Leading Emerg- ing Sciences Conference (NILES), IEEE, 2023, pp. 161–165. doi: 10.1109/NILES59815.2023.10296582

work page doi:10.1109/niles59815.2023.10296582 2023

[4] [4]

Explaining a deep learning based breast ultrasound image classifier with saliency maps,

M. Byra, K. Dobruch-Sobczak, H. Piotrzkowska-Wroblewska, Z. Klimonda, and J. Litniewski, “Explaining a deep learning based breast ultrasound image classifier with saliency maps,” Journal of Ultrasonography, vol. 22, pp. e70–e75, 2022. doi: 10.15557/JoU.2022.0013

work page doi:10.15557/jou.2022.0013 2022

[5] [5]

Breast UltraSound Image classification using fuzzy-rank-based ensemble network,

S. D. Deb and R. K. Jha, “Breast UltraSound Image classification using fuzzy-rank-based ensemble network,”Biomedical Signal Processing and Control, vol. 85, p. 104871, 2023

work page 2023

[6] [6]

Vision transformers for classification of breast ultrasound images,

B. Gheflati and H. Rivaz, “Vision transformers for classification of breast ultrasound images,” inProc. 44th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), pp. 480– 483, 2022

work page 2022

[7] [7]

Deep learning approaches for classification of breast cancer in ultrasound (US) images,

I. Pacal, “Deep learning approaches for classification of breast cancer in ultrasound (US) images,”Journal of the Institute of Science and Tech- nology, vol. 12, no. 4, pp. 1917–1927, 2022. doi: 10.21597/jist.1183679

work page doi:10.21597/jist.1183679 1917

[8] [8]

Breast cancer classification based on convolutional neural network and image fusion approaches using ultrasound images,

M. Alotaibi, A. Aljouie, N. Alluhaidan, W. Qureshi, H. Almatar, R. Al- duhayan, B. Alsomaie, and A. Almazroa, “Breast cancer classification based on convolutional neural network and image fusion approaches using ultrasound images,”Heliyon, vol. 9, no. 11, p. e22406, 2023. doi: 10.1016/j.heliyon.2023.e22406

work page doi:10.1016/j.heliyon.2023.e22406 2023

[9] [9]

Enhancing breast cancer segmentation and classification: An ensemble deep convolu- tional neural network and U-net approach on ultrasound images,

M. R. Islam, M. M. Rahman, M. S. Ali, A. A. N. Nafi, M. S. Alam, T. K. Godder, M. S. Miah, and M. K. Islam, “Enhancing breast cancer segmentation and classification: An ensemble deep convolu- tional neural network and U-net approach on ultrasound images,” Machine Learning with Applications, vol. 16, p. 100555, 2024. doi: 10.1016/j.mlwa.2024.100555

work page doi:10.1016/j.mlwa.2024.100555 2024

[10] [10]

An EfficientNet integrated ResNet deep network and explainable AI for breast lesion classification from ultrasound images,

K. Jabeen, M. A. Khan, A. Hamza, H. M. Albarakati, S. Alsenan, U. Tariq, and I. Ofori, “An EfficientNet integrated ResNet deep network and explainable AI for breast lesion classification from ultrasound images,”CAAI Transactions on Intelligence Technology, vol. 10, no. 3, pp. 842–857, 2025. doi: 10.1049/cit2.12385

work page doi:10.1049/cit2.12385 2025

[11] [11]

Improving breast cancer diagnosis in ultrasound images using deep learning with feature fusion and attention mechanism,

S. Asif, Y . Yan, B. Feng, M. Wang, Y . Zheng, T. Jiang, R. Fu, J. Yao, L. Lv, M. Song, L. Sui, Z. Yin, V . Y . Wang, and D. Xu, “Improving breast cancer diagnosis in ultrasound images using deep learning with feature fusion and attention mechanism,”Academic Radiology, vol. 32, pp. 4997–5009, 2025. doi: 10.1016/j.acra.2025.05.007

work page doi:10.1016/j.acra.2025.05.007 2025

[12] [12]

Breast can- cer classification from ultrasound images using VGG16 model based transfer learning,

A. B. M. A. Hossain, J. K. Nisha, and F. Johora, “Breast can- cer classification from ultrasound images using VGG16 model based transfer learning,”International Journal of Image, Graphics and Sig- nal Processing (IJIGSP), vol. 15, no. 1, pp. 12–22, 2023. doi: 10.5815/ijigsp.2023.01.02

work page doi:10.5815/ijigsp.2023.01.02 2023

[13] [13]

Advancing breast ultrasound diagnostics through hybrid deep learning models,

A. Kiran, J. V . N. Ramesh, I. S. Rahat, M. A. U. Khan, A. Hossain, and R. Uddin, “Advancing breast ultrasound diagnostics through hybrid deep learning models,”Computers in Biology and Medicine, vol. 180, p. 108962, 2024. doi: 10.1016/j.compbiomed.2024.108962

work page doi:10.1016/j.compbiomed.2024.108962 2024

[14] [14]

An explorative analysis of SVM classifier and ResNet50 architecture on African food classification,

C. Mbonu, K. Anigbogu, D. Asogwa, and T. Belonwu, “An explorative analysis of SVM classifier and ResNet50 architecture on African food classification,”arXiv preprint arXiv:2505.13923, 2025

work page arXiv 2025

[15] [15]

EfficientNet: Rethinking model scaling for convolutional neural networks,

M. Tan and Q. V . Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” inProc. 36th International Conference on Machine Learning (ICML), vol. 97, pp. 6105–6114, 2019

work page 2019

[16] [16]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, “Attention is all you need,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, pp. 5998– 6008, 2017

work page 2017

[17] [17]

Focal loss for dense object detection,

T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” inProc. IEEE International Conference on Computer Vision (ICCV), pp. 2980–2988, 2017

work page 2017

[18] [18]

Decoupled Weight Decay Regularization

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inarXiv preprint arXiv:1711.05101, 2017

work page internal anchor Pith review Pith/arXiv arXiv 2017

[19] [19]

A review on evaluation metrics for data classification evaluations,

M. Hossin and M. N. Sulaiman, “A review on evaluation metrics for data classification evaluations,”International Journal of Data Mining and Knowledge Management Process, vol. 5, no. 2, pp. 1–11, 2015

work page 2015

[20] [20]

Metrics for multi-class classifi- cation: an overview,

M. Grandini, E. Bagli, and G. Visani, “Metrics for multi-class classifi- cation: an overview,”arXiv preprint arXiv:2008.05756, 2020

work page arXiv 2008

[21] [21]

An introduction to ROC analysis,

T. Fawcett, “An introduction to ROC analysis,”Pattern Recognition Letters, vol. 27, no. 8, pp. 861–874, 2006

work page 2006

[22] [22]

A multi-algorithmic approach to stroke risk prediction using machine learning,

E. G. Onyedinma, D. C. Asogwa, T. S. Belonwu, and C. E. Mbonu, “A multi-algorithmic approach to stroke risk prediction using machine learning,”Journal of Engineering Research and Reports, vol. 27, no. 7, pp. 247–259, 2025. doi: 10.9734/jerr/2025/v27i71573

work page doi:10.9734/jerr/2025/v27i71573 2025