Multi-Scale Curriculum Learning for Efficient Automatic Whole Slide Image Segmentation

Abstract

Deep learning becomes a powerful tool for multiple tasks in computational pathology and has achieved remarkable performance in automatic tumor segmentation of whole slide images (WSIs). Curriculum learning is an effective learning strategy that trains a deep neural network from easy to hard samples and has been applied to the problem of tumor segmentations in WSIs. However, it requires measuring the "difficulty" of histopathological images by pathologists, which is challenging. Here we propose a curriculum learning strategy that does not require additional difficulty labeling based on the conjecture that high resolution labeling is more challenging than low resolution labeling. Our multi-scale curriculum learning strategy allows us to control the difficulty of segmentation labels for more efficient training. Our experiments on the PAIP prostate cancer dataset validate the effectiveness of our proposed multi-scale curriculum learning strategy, showing improved performance in tumor segmentation qualitatively and quantitatively.