2020/02/20 by Maxime W. Lafarge, Lafarge, Maxime W., Erik J. Bekkers +7 · 3 citations
Computer Science · Engineering · #AI in cancer detection #Computer Vision and Pattern Recognition (cs.CV) #Digital Imaging for Blood Diseases #FOS: Computer and information sciences #Medical Imaging and Analysis
paper · pdf · doi:10.48550/arxiv.2002.08725
openalex publication_date 2020/02/20 · openalex created_date 2022/07/26 · openalex updated_date 2026/07/28
Rotation-invariance is a desired property of machine-learning models for\nmedical image analysis and in particular for computational pathology\napplications. We propose a framework to encode the geometric structure of the\nspecial Euclidean motion group SE(2) in convolutional networks to yield\ntranslation and rotation equivariance via the introduction of SE(2)-group\nconvolution layers. This structure enables models to learn feature\nrepresentations with a discretized orientation dimension that guarantees that\ntheir outputs are invariant under a discrete set of rotations. Conventional\napproaches for rotation invariance rely mostly on data augmentation, but this\ndoes not guarantee the robustness of the output when the input is rotated. At\nthat, trained conventional CNNs may require test-time rotation augmentation to\nreach their full capability. This study is focused on histopathology image\nanalysis applications for which it is desirable that the arbitrary global\norientation information of the imaged tissues is not captured by the machine\nlearning models. The proposed framework is evaluated on three different\nhistopathology image analysis tasks (mitosis detection, nuclei segmentation and\ntumor classification). We present a comparative analysis for each problem and\nshow that consistent increase of performances can be achieved when using the\nproposed framework.\n