Representative Region Based Active Learning For Histological Classification Of Colorectal Cancer.

The advent of advanced deep learning algorithms has contributed to many successful applications in digital histopathology. Regularly, attributed to the extremely large size, whole slide images (WSIs) have to be decoupled into numerous smaller patches to be independently processed by convolutional neural networks (CNNs) for training and inference. The routine tessellation strategy chops the images by a sliding-window moving across the entire whole slide. The straightforward result of this methodology is a large proportion of non-representative patches, as well as a labor-intensive manual annotation. Although active learning based models to sort out informative data instances are many, the identification of the most informative region of patches to train a patch-wise classifier has not been discussed. In this research, we propose an active learning based model to select patches with optimized offsets in and spatial adaptive manner. With these most representative patches, the patch-level classification models can be more effectively and efficiently trained. To the best of our knowledge, this is the first literature on an adaptive representative patch generation system. The empirical results on large patient cohorts in The Cancer Genome Atlas (TCGA) show a scale reduction in training set of 38.0% can achieve the tumor classification accuracy of 92.70%.