The document proposes a relevance-based compression method for cataract surgery videos using convolutional neural networks. The method uses Mask R-CNN to detect relevant regions like the cornea and instruments. Pixels outside these regions are removed or compressed at lower quality. Testing showed the method achieved up to 68% reduction in video size while maintaining good quality for relevant regions. The summaries provide the key information about the proposed method and results at a high level in 3 sentences or less as requested.