Abstract 2098: AzinNet: A wavelet convolutional neural network for pathology image analysis

Abstract

Abstract Deep learning models are gaining popularity for pathology image analysis in cancer and have successfully been used to build disease and mutation status prediction models [1,2]. However, there are several shortcomings to these models. They typically tile whole slide images with little or no pre-processing leaving small artifacts, such as stitching artifacts, intact in the training data. The network may correlate these artifacts with class labels instead of, or in addition to, biologically meaningful morphological features. While one may argue transfer learning can be used to guard against such artifacts, (1) to the best of our knowledge no rigorous study has been performed to validate this idea given whole slide images, and (2) transfer learning-based deep learning models, such as the one studied in [2], show sensitivity to small artifacts in the data. That being said, transfer learning based methods usually borrow from models trained on the image-net dataset, in which each image contains one object we wish to classify. In contrast, pathology images contain morphological features at different scales spread through the slide, which may be best understood as texture-type features. This begs the question of whether transfer learning models, when used as black-box universal feature extractors, miss important morphological features due to the different nature of features encountered in pathology images. To address these issues we propose a wavelet based convolutional neural network, called AzinNet, inspired by the wavelet based network of [3]. AzinNet uses wavelet transform to (1) denoise the data and remove small artifacts, and (2) create a hierarchy of wavelet channels representing morphological features at different scales. We train and test AzinNet on TCGA whole slide images, and compare it with transfer learning based models inputted with raw and denoised images. We also compare robustness of AzinNet and transfer learning-based models to small perturbations of the data. Initial results suggest AzinNet achieves a higher area under curve and is more robust to data perturbations.