Annotation-Free Deep Learning for Predicting Gene Mutations from Whole Slide Images of Acute Myeloid Leukemia

Abstract

The rapid development of deep learning in recent years has revolutionized the field of medical image processing, including the applications of using high-resolution whole slide images (WSIs) in acute myeloid leukemia (AML) diagnosis. Although the potential of characterizing gene mutations directly from WSIs has been demonstrated in some cancers, it still faces challenges due to image resolutions and manual annotations. To address this, we propose a deep learning model based on multiple instance learning (MIL) with ensemble learning to predict gene mutations from AML annotation-free WSIs. Our deep learning model offers a promising solution for gene mutation prediction on NPM1 mutations and FLT3 -ITD without the need for patch-level or cell-level manual annotations, reducing the manpower and time costs associated with traditional supervised learning approaches. The dataset of 572 WSIs from AML patients that we used to train our MIL models is currently the largest independent database with both WSI and genetic mutation information. By leveraging upsampling and ensemble learning techniques, our final model achieved an AUC of 0.90 for predicting NPM1 mutations and 0.81 for FLT3 -ITD. This confirms the feasibility of directly obtaining gene mutation data through WSIs without the need for expert annotation and training involvement. Our study also compared the proportional representation of cell types before and after applying the MIL model, finding that blasts are consistently important indicators for gene mutation predictions, with their proportion increasing in mutated WSIs and decreasing in non-mutated WSIs after MIL application. These enhancements, leading to more precise predictions, have brought AML WSI analysis one step closer to being utilized in clinical practice.