Reproduce “Decompose-and-Integrate Learning for Multi-class Segmentation in Medical Images” on OpenResearch

About this paper

Yizhe Zhang, Michael T. C. Ying, Danny Z. Chen

The research introduces "Decompose-and-Integrate Learning," a novel scheme for multi-class medical image segmentation that strategically preprocesses annotation maps to guide deep networks toward learning richer, more disentangled feature representations. This approach consistently improves segmentation performance, achieving state-of-the-art results on datasets like HVSMR for cardiovascular structures and GlaS for gland segmentation, with performance gains attributed to the explicit decomposition of learning tasks.

Key insights

• Strategically decomposing complex multi-class segmentation into simpler sub-problems, each with its dedicated learning pathway, encourages deep networks to learn richer and more disentangled feature representations.
• Explicitly guiding the network to focus on specific object classes, shapes, or image-level contexts during an initial decomposition phase prevents overshadowing of difficult classes and improves the distinctness of learned features.
• An end-to-end trainable K-to-1 network can effectively integrate the specialized knowledge gained from solving decomposed sub-problems, leading to a superior overall solution for the original multi-class segmentation task.

Summary by alphaXiv.