Two-Stage Deep Learning Segmentation for Tiny Brain Regions

Accurate segmentation of brain regions has become increasingly important in the early diagnosis of brain diseases. Widely used methods for brain region segmentation usually rely on atlases and deformations, which require manual intervention and do not focus on tiny object segmentation. To address the challenge of tiny brain regions segmentation, we propose a two-stage segmentation network based on deep learning, using both 2D and 3D convolution. We first introduce the concept of the Small Object Distribution Map (SODM), allowing the model to perform coarse-to-fine segmentation for objects of different scales. Then, a contrastive loss function is implemented to automatically mine difficult negative samples, and two attention modules are added to assist in the accurate generation of the small object distribution map. Experimental results on a dataset of 120 brain MRI demonstrate that our method outperforms existing approaches in terms of objective evaluation metrics and subjective visual effects and shows promising potential for assisting in the diagnosis of brain diseases.