Predicting genetic alterations in Acute Myeloid Leukemia through AI-based analysis of bone marrow morphology

Abstract Acute myeloid leukemia (AML) is characterized by significant genetic heterogeneity, with diverse cytogenetic and molecular alterations that profoundly influence prognosis and therapeutic decisions. Rapid and precise identification of key alterations such as PML::RARA, CBFB::MYH11, RUNX1::RUNX1T1, NPM1, and TP53 is crucial for timely diagnosis, accurate prognostication, and optimal therapeutic intervention. However, conventional cytogenetic and molecular diagnostics may require considerable time and specialized facilities, limiting their immediate availability. Artificial intelligence (AI)-assisted morphological analysis presents a potential solution to rapidly and reliably predict the presence of these genetic alterations through routine morphological assessment of bone marrow samples. Molecular or cytogenetic data for each genetic alteration were available in a subset of the total 287 AML patients analyzed, resulting in varying sample sizes across alterations. Samples analyzed included subsets positive for specific genetic alterations: PML::RARA (287 samples, 15 positives), CBFB::MYH11 (154 samples, 7 positives), RUNX1::RUNX1T1 (156 samples, 9 positives), NPM1 (up to 254 samples, 67 positives), and TP53 (234 samples, 39 positives). Digital images were acquired using a standard optical microscope coupled to a mobile phone through a custom 3D-printed adapter, digitizing multiple microscopic fields per patient at 100x magnification. An initial AI algorithm, trained on more than 400,000 manually labeled cells, automatically classified nucleated cells into 23 distinct cell types with high precision. Morphological features were then extracted from each cell using a foundational AI model pretrained on a database of more than 6 million unlabeled cells. Subsequently, multiple-instance learning (MIL) models were independently trained and evaluated for every possible combination of genetic alteration and cell type (or groups of cell types). Combinations yielding the best predictive performance for each genetic alteration were identified and selected for clinical evaluation. Each selected model was validated using five-fold cross-validation, maintaining strict patient-level separation between training and validation sets. All models were evaluated using the area under the ROC curve (ROC-AUC), the area under the precision-recall curve (PR-AUC), sensitivity (Sens), specificity (Spec), positive predictive value (PPV), and negative predictive value (NPV). The AI-based MIL models demonstrated robust predictive performance across all studied genetic alterations. For PML::RARA: ROC-AUC 0.999 ± 0.003, PR-AUC 0.983 ± 0.037, Sens 93.3% ± 14.9%, Spec 99.3% ± 1.0%, PPV 90.0% ± 13.7%, NPV 99.6% ± 0.8%. For CBFB::MYH11: ROC-AUC 0.928 ± 0.125, PR-AUC 0.761 ± 0.238, Sens 70% ± 44.7, Spec 94.5% ± 5.2, PPV 34% ± 23%, NPV 98.6% ± 1.9%. For RUNX1::RUNX1T1: ROC-AUC 0.956 ± 0.042, PR-AUC 0.73 ± 0.30, Sens 80% ± 27.4, Spec 96% ± 6, PPV 70.7% ± 40.4, NPV 98.6% ± 1.9. For NPM1: ROC-AUC 0.83 ± 0.03, PR-AUC 0.66 ± 0.09, Sens 82.8% ± 6.4, Spec 77.7% ± 7.8, PPV 59.8% ± 7.7, NPV 92.3% ± 2.4. For TP53: ROC-AUC 0.758 ± 0.061, PR-AUC 0.567 ± 0.078, Sens 53.9% ± 10.5, Spec 89.2% ± 6.9, PPV 53.5% ± 11.4, NPV 90.7% ± 1.7. AI models analyzing morphological features from microscopy images provide promising accuracy for predicting certain genetic alterations in AML, notably PML::RARA and RUNX1::RUNX1T1, potentially enhancing diagnostic workflows, while alterations such as TP53 remain more challenging to predict based solely on morphology. This study establishes a foundation for extending research to larger patient cohorts to enhance prediction based on digital imaging.