DeepPyramid: Enabling Pyramid View and Deformable Pyramid Reception for Semantic Segmentation in Cataract Surgery Videos

Abstract Semantic segmentation in cataract surgery has a wide range of applications contributing to surgical outcome enhancement and clinical risk reduction. However, the varying issues segmenting different relevant structures these surgeries make designation unique network quite challenging. This paper proposes semantic network, termed DeepPyramid, that can deal with challenges using three novelties: (1) Pyramid View Fusion module which provides varying-angle global view surrounding region centering at each pixel position input convolutional feature map; (2) Deformable Reception enables deformable receptive field adapt geometric transformations object interest; (3) dedicated Loss adaptively supervises multi-scale maps. Combined, we show modules effectively boost performance, especially case transparency, deformability, scalability, blunt edges objects. We demonstrate our approach performs state-of-the-art level outperforms number existing methods large margin (\(3.66\%\) overall improvement intersection over union compared best rival approach).KeywordsCataract surgerySemantic segmentationSurgical data science