New DACMC loss function boosts medical segmentation with curvature constraints

Medical image segmentation is critical for diagnosis, but deep learning models often miss geometric priors because they train at the pixel level. Existing methods like Chan-Vese loss account for region boundaries but fail to capture fine curvature details. To solve this, Xiao-qiang Zhai and colleagues introduce the DACMC loss function, which incorporates mean curvature as a natural geometric constraint. They approximate the mean curvature using a convolution kernel, making the computation efficient enough for training.

Tested on liver and spleen segmentation datasets, DACMC outperforms prior approaches, setting new state-of-the-art results. The curvature term helps the model focus on smooth, anatomically plausible boundaries. This method is particularly promising for clinical workflows where precise organ delineation is essential. The paper is under review at Biomedical Signal Processing and Control.