We propose an end-to-end unsupervised deep learning registration model consisting of an affine and a deformable module to spatially align histopathology images. The learned Displacement Vector Field is provided to a spatial transformer network to generate registered image. The superior performance of our proposed model to other methods suggests its promising potential for IHC histopathology image registration.
The mortality rate due to brain cancer is the highest in the Asian continent according to WHO report. Glioblastoma (GBM) are malignant Grade IV tumors an aggressive type of cancer. Since cancer diagnosis is high invasive, expensive and time consuming, it is essential to develop a non-invasive, cost-effective and efficient tools for brain cancer characterization and grade estimation. In this study we propose a Recurrent Residual Unet based segmentation framework with multi-resolution feature for semantic regions segmentation such as necrosis, tumor, stroma etc. in Whole-slide histopathology images of breast cancer(Norway) and brain tumor data(TCGA and IVYGlioblastoma). The quantitative and qualitative results suggest the superior performance of our proposed model compared to the state-of-the-art deep learning segmentation models.
In this study immunohistochemistry (ISH) stained breast cancer images are used for feature difference analysis of Aromatase expressing positive vs negative tumor cells. First, we separate H and DAB color channels followed by segmenting tumor regions in each small image patch. Once we have segmented tumor regions in H channel, the existence of corresponding instances in DAB color channel can be used to find the Aromatase expression positive vs negative tumor cells. We used various supervised, and semi supervised deep learning-based models for tumor regions (isolated or clumped with weak separating borders) segmentation along with their performance comparison.