Automatic Kidney Segmentation

Automatically segmenting kidneys in ultrasound images is challenging because kidney shape and image intensity vary widely, limiting the success of fully automated methods. The study introduces a two-step deep learning approach that first predicts kidney boundary distance maps using features from pre-trained neural networks, then classifies pixels as kidney or non-kidney in an end-to-end framework. With additional data augmentation based on kidney shape registration, this method outperformed standard pixel classification networks and achieved strong segmentation performance from a small labeled dataset.