Improving medical image segmentation using contour-weighted loss

Motivation: Image segmentation holds significant importance in medical image analyses. Nevertheless, accurately segmenting images presents challenges, particularly due to data imbalances that are often encountered in multi-targets segmentation. Goal(s): We aim to develop a model-independent loss function to enhance the multi-targets segmentation of medical images . Approach: We develop a loss function, which integrates a contour-weighted cross-entropy loss with a separable dice loss. Moreover, we design a partial decoder attention network, named PDANet, to refine segmentation performance. Results: Results on the BraTS dataset reveal that our loss function surpassed other existing methods, improving segmentation accuracy in widely used models, with our approach achieving superior results. Impact: We developed a contour-weighted loss function to address the problem of data imbalance in medical image segmentation. Our approach is model-independent, allowing it to integrate seamlessly with any segmentation network, thereby improving segmentation performance across different models.