Advancing Lightweight Breast Cancer Detection Models
Open Access DepositedParameter-Efficient Fine-Tuning and Knowledge Distillation from Foundational Models
Breast cancer remains one of the leading causes of cancer-related mortality among women, with disproportionately rising incidence rates in underdeveloped regions. While early detection through medical imaging can significantly improve outcomes, access to advanced diagnostic tools is often limited by infrastructure and resource constraints. Recent advancements in vision foundation models, such as Segment Anything Model 2 (SAM2) and MedSAM, offer strong segmentation capabilities but are computationally intensive. This study investigates multiple strategies to improve the efficiency and accessibility of breast cancer segmentation models without sacrificing performance. Parameter-efficient fine-tuning (PEFT) was explored using both SAM2 and MedSAM, applied across multiple augmented versions of a single mammography dataset to evaluate model robustness under different image enhancement strategies. The base SAM2 model demonstrated consistently strong and stable performance—outperforming MedSAM despite the latter’s domain-specific medical pretraining—suggesting that architectural advancements in SAM2 offer greater benefit for this task. Building on this foundation, knowledge distillation (KD) was employed to train lightweight student models using both convolutional (ResNet) and transformer-based (ViT) architectures. Results show that distilled ViT-based student models not only outperformed their counterparts trained from scratch, but also matched—and in some cases exceeded—the segmentation performance of their teacher model, despite being as small as 4% of the original model size. Notably, this research achieves these results without the use of any spatial or text-based prompts in either the teacher or student models—an uncommon approach given that SAM-family models typically rely on such prompting for effective segmentation. These findings demonstrate the potential of KD to enable high-performing, efficient breast cancer detection models that are suitable for deployment in a wide range of clinical settings, including those with limited computational resources.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.