Deep learning predicts IDH1 mutation status in Acute Myeloid Leukemia from bone marrow smear images

Abstract

Abstract Acute myeloid leukemia (AML) is a genetically heterogeneous disease. The rise of next-generation-sequencing (NGS) and its incorporation in the clinical routine has unveiled numerous recurrently altered genes in AML biology, a minority of which are also actionable via targeted drugs. Alterations of the isocitrate dehydrogenase 1 (IDH1) are recurrent events in AML occurring in approximately 8% of patients. These mutations impair normal cellular metabolism leading to an accumulation of the oncometabolite 2-hydroxyglutarate (2-HG), affecting epigenetic regulation and cellular differentiation. Ivosidenib is a selective oral inhibitor of mutant IDH1, reducing the abnormal accumulation of 2-HG and has been approved by the FDA or AML both in the upfront and relapsed/refractory setting. Recent clinical trials suggest improved outcomes for ivosidenib-based combination therapies compared to non-ivosidenib-based combinations in IDH1-mutated AML patients ineligible for intensive treatment. Hence, identifying all IDH1-mutated cases is prognostically relevant. Deep Learning (DL) has been demonstrated to uncover genotype-phenotype links in bone marrow cytomorphology. Pre-treatment samples from initial diagnosis were selected retrospectively from AML patients treated at University Hospital Dresden, Germany. May-Gruenwald-Giemsa stained bone marrow smears (BMS) were digitized to generate whole slide images (WSIs) of 86 AML patients with alterations of IDH1 and 60 AML patients with IDH1 wildtype as confirmed by NGS. 40 tiles at 1263*1263 pixels were selected from each WSI. The resulting 5840 images were allocated to training and test set at an 80:20 ratio, ensuring that no individual patient contributed images to both sets. We evaluated eight state-of-the-art convolutional neural networks (Resnet, Resnext, Shufflenet, Squeezenet, Densenet, Efficientnet, Regnet, and Convnext), performing a hyperparameter search for each network. Tile extraction, model training, and testing was performed at the Center for Information Services and High Performance Computing (ZIH), Dresden University of Technology with eight NVIDIA A100-SXM4 tensor core graphical processing units. All models were evaluated on a held-out patient-level test set. For the binary classification of image-based IDH1 mutation status prediction, we observed an area-under-the-receiver-operating-characteristic (AUROC) ranging from 0.731 to 0.788 with a corresponding accuracy ranging from 0.728 to 0.774 and an F1-score ranging from 0.676 to 0.741. The generation of digital biomarkers predicting image-based mutation status of targetable genetic alterations such as IDH1 allows for image-based genotypic screenings as currently not all AML patients are appropriately screened for targetable alterations. Hence, the incorporation of digital tools into routine bone marrow microscopy ensures a quick pre-screening and may allow for a broader use of targeted therapy to improve outcomes.