PathTR: Context-Aware Memory Transformer for Tumor Localization in Gigapixel Pathology Images
With the development of deep learning and computation pathology, whole-slide images (WSIs) are wildly used in clinical diagnosis. The WSI, which refers to the scanning of conventional glass slides in order to produce digital slides, usually has gigapixels. Most existing methods in computer vision process WSIs as many patches. The model infers patch by patch to get the results on WSI, which loses the global context of WSI. In this paper, we developed PATHology TRansformer (PathTR), which fully uses the global information of WSI. In PathTR, the local context is aggregated by the self-attention mechanism. We further design a recursive mechanism to encode the global context in extra states. In tumor detection of metastases of lymph node sections for breast cancer, we got the FROC score of 87.68% which outperforms the baseline and NCRF method with +8.99% and +7.08%, respectively. We highlight that we also achieve a significant 94.25% sensitivity at 8 false positives per image.