Breast Cancer Ultrasound CAD: Sequential vs Multi-task Deep Learning
BIT Genesis Research Award 2025 @ UEH | Presented at NCTD 2025 National Conference
Faculty-level research comparing Sequential and Multi-task Learning architectures for breast cancer diagnosis from ultrasound images. Built on U-Net with EfficientNet-B4 backbone, systematically evaluated Deformable Convolution and Capsule Network modules through ablation study. Evaluated on BUSI dataset (780 images: Normal/Benign/Malignant) with rigorous statistical testing (Shapiro-Wilk, Mann-Whitney U, Kruskal-Wallis, Tukey HSD).
Timeline
2025
Type
Research
Status
completed
Outcome / Impact
- •BIT Genesis Research Award 2025 at Business Information Technology Department, UEH
- •Paper presented at National Conference on Technology and Design 2025 (NCTD 2025) – Shaping Vietnam's Digital Future
- •MTL achieved statistically significant superiority in classification (Accuracy 85.26% vs 62.05%, p<0.05) with Mann-Whitney U test
- •No significant difference in segmentation (Dice ~76%) between architectures — both approaches equally effective
- •Ablation study revealed: advanced modules (Deformable Conv, CapsNet) did NOT improve performance vs traditional Conv+FC
- •Established deployment trade-off: MTL for hospital CAD (accuracy priority) vs Sequential for rapid screening (speed priority)
Tech / Skills
Certificates (2)
Case Study
1) Context / Problem
Breast cancer is the most common cancer in women globally. Ultrasound is a cost-effective, non-invasive imaging modality but interpretation depends heavily on operator expertise. Existing studies focused on either Sequential (two-stage) or Multi-task Learning architectures without direct, controlled comparison on the same baseline and dataset.
2) Your Role
As team member, I was responsible for data preprocessing on BUSI dataset (780 ultrasound images with masks), implementing model architectures (U-Net with EfficientNet-B4, Deformable ConvBlocks, Capsule Network), conducting ablation experiments, statistical hypothesis testing, and co-authoring the research paper.
3) Approach
Built unified evaluation framework with shared U-Net + EfficientNet-B4 backbone. Created 8 model variants combining: (1) Architecture type (Sequential vs MTL), (2) Conv type (Standard vs Deformable), (3) Classification head (FC vs CapsNet). Used Uncertainty-Weighted Loss for MTL. Applied rigorous statistical pipeline: Shapiro-Wilk → Levene's → Mann-Whitney U / Kruskal-Wallis → Dunn's / Tukey HSD post-hoc tests.
4) Result / Impact
Won BIT Genesis Research Award 2025 and paper accepted for presentation at NCTD 2025 National Conference. MTL architecture significantly outperformed Sequential in classification (85.26% vs 62.05% accuracy, p<0.05). No statistically significant difference in segmentation performance. Surprisingly, advanced modules (Deformable Conv, CapsNet) showed NO improvement or even degraded performance compared to standard Conv+FC.
5) Learnings
Multi-task learning enables effective feature sharing between segmentation and classification tasks. Simpler architectures often outperform complex ones in constrained data settings. Rigorous statistical testing is essential — visual differences in metrics can be misleading without hypothesis testing. Future work: larger datasets, cross-hospital validation, multimodal integration.
6) Links
See links above.