A Multi-task Learning Framework for Carotid Plaque Area Measurement in Imbalanced Datasets

The measurement of carotid plaque area is of critical clinical significance for atherosclerosis risk assessment and stroke prediction. However, existing segmentation methods exhibit limitations in both the accuracy and robustness of plaque area prediction. To address this issue, this paper proposes an innovative multi-task learning framework that enhances primary segmentation through edge segmentation and simultaneously enables direct regression for plaque area prediction. The proposed framework integrates a shared encoder, dual decoders (segmentation and edge decoders), and an area regression sub-network, collaboratively optimizing segmentation and plaque area detection tasks. Additionally, a dynamic weighting strategy is introduced to tackle data imbalance, improving model stability and accuracy. Experimental results demonstrate that the proposed method significantly outperforms existing techniques across two independent datasets, particularly in plaque area prediction and segmentation accuracy. This method provides an efficient and reliable solution for plaque segmentation and quantitative analysis in carotid ultrasound images, showing substantial clinical potential for atherosclerosis risk assessment and stroke prediction.