Team: The Warwick-QU Team, Warwickshire, UK
Authors: Muhammad Shaban, Talha Qaiser, Ruqayya Awan, Korsuk Sirinukunwattana, Yee-Wah Tsang, and Nasir Rajpoot
Our approach aims at segmenting the tumor regions by using a variant of the U-Net convolutional-deconvolutional neural network as the main component. Before feeding image patches to the network, we first performed background removal and stain normalization to deal with stain variation in the H&E stained slide images. We segmented the tissue region (area of interest) from fatty and white background areas by using a simple fully convolutional network (FCN) with a single upsampling layer. We then implemented the architecture of U-Net that is customised for the task of tumor segmentation in many ways. Our implementation in TensorFlow allows ease of customisation. The output of the network is a probability map, whereby each pixel intensity value corresponds to the probability of it belonging to the tumor, which is further manipulated by taking into account the probability values within a region and the region area. The candidate regions are further refined using the weighted probability map. For training purposes, we collected 20,000 RGB patches of size 428X428 at magnification level 2 (10X), with 12,000 patches taken from normal WSIs and 8,000 patches from WSIs with metastasis. The training and validation datasets comprised of 90% and 10% of the whole dataset respectively and the network was trained for more than 50 epochs.
The following figure shows the receiver operating characteristic (ROC) curve of the method.
The following figure shows the free-response receiver operating characteristic (FROC) curve of the method.
The table below presents the average sensitivity of the developed system at 6 predefined false positive rates: 1/4, 1/2, 1, 2, 4, and 8 FPs per whole slide image.
Error rendering graph from file