Local image fitting based active contour loss with deep learning for nuclei segmentation

This paper proposes an approach for segmentation of nuclei images based on deep learning. In particular, the recent TransUnet inspired from transformers’ strong ability in modeling long-range context, is employed and adapted for the nuclei segmentation. For training the neural network, we propose a new loss inspired from active contour models with the guide of local image fitting. The loss when applied for the TransUnet has shown promising results over common Dice and Binary Cross Entropy loss functions. Our approach has been validated on the Data Science Bowl 2018 dataset, which includes 670 data folders for training model and 65 data folders for testing. State of the art models, such as FCN, SegNet, Unet, and DoubleU-Net are also conducted and evaluated. Quantitative assessments with high Dice similarity coefficient and Intersection over Union metrics demonstrate the performances of the proposed approach for nuclei segmentation.