A Coarse-to-Fine Pathology Patch Selection for Improving Gene Mutation Prediction in Acute Myeloid Leukemia

Identifying gene mutation is essential to prognosis and therapeutic decisions for acute myeloid leukemia (AML) but the current gene analysis is inefficient and non-scalable. Pathological images are readily accessible and can be effectively modeled using deep learning. This work aims at predicting gene mutation directly by modeling bone marrow smear images. Traditionally, bone marrow smear slides are cropped into patches with manual segmentation for patch-level modeling. Slide-level modeling, such as multi-instance learning, could aggregate patches for holistic modeling, though suffer from excessive redundancy. In this study, we propose a discriminative multi-instance approach to select useful patches in a coarse-to-fine process. Specifically, we preprocess a slide into patches by using a trained pre-selector network. Then, we rule out low quality patches in the coarse selection with known prior knowledge, and refine the model using gene-discriminative patches in the fine selection. We evaluate the framework for CEBPA, FLT3, and NPM1 gene mutation prediction and obtain 71.67%, 56.26%, and 56.34% F1-score. Further analysis show the effect of different selection criteria on prediction gene mutations using pathological images.