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 predicting gene mutations in acute myeloid leukemia (AML). Although the potential of characterizing gene mutations directly from WSIs has been demonstrated in some studies, it still faces challenges due to memory limitations and manual annotation requirements. To address this, we propose a deep learning model based on multiple instance learning (MIL) to predict gene mutations from AML WSIs with no patch-level or cell-level annotations. The proposed MIL-based deep learning model offers a promising solution for gene mutation prediction on NPM1 mutations and FLT3 -ITD. With the property of annotation-free, the proposed method eliminates the need for manual annotations, reducing the manpower and time costs associated with traditional patch-based or cell-based approaches. We assessed our MIL models using a dataset of 572 WSIs from AML patients. By exclusively utilizing annotation-free WSIs for cell-level training, we achieved an AUC of 0.75 for predicting NPM1 mutations and 0.68 for FLT3 -ITD. Furthermore, upon applying upsampling and ensemble techniques to address the data imbalance issue, the AUC improved from 0.75 to 0.86 for NPM1 mutations and from 0.68 to 0.82 for FLT3 -ITD. These enhancements, leading to more precise predictions, have brought AML WSI analysis one step closer to being utilized in clinical practice.