Summary
This study addresses the challenges of weakly supervised segmentation in medical images, particularly in the context of thyroid ultrasound images. The variations in nodule sizes and the limitations of existing methods lead to segmentation problems such as under and over-segmentation.
To overcome these issues, the researchers propose a novel neural network approach. It incorporates a dual branch soft erase module to enhance the foreground response region while controlling erroneous expansion, a scale feature adaptation module to improve sensitivity to nodule size, and an edge-based attention mechanism for better nodule edge segmentation.
Experimental results on thyroid ultrasound images demonstrate that this approach outperforms existing weakly supervised segmentation methods, achieving higher accuracy in terms of Jaccard and Dice coefficients.
