Abstract 466: Predicting survival of rhabdomyosarcoma patients based on deep learning of H&E images

Abstract

Abstract Background: Rhabdomyosarcoma (RMS) is an aggressive soft tissue tumor in children and young adults, accounting for 350-400 new cases annually in the US. Diagnosis of RMS is defined by the expression of genes related to skeletal muscle differentiation and can be further subclassified based on histological patterns (embryonal, ERMS; alveolar, ARMS; spindle/sclerosing, SSRMS). Genetic studies have found that the presence of a PAX fusion gene (FP-RMS), which is present in many ARMS tumors, correlates with poor outcome. Subsequent studies have identified additional genetic alterations (ex. TP53 or MYOD1 mutations) which also display distinct histological features and are associated with poor outcome. As a result, there is a growing need to identify these mutations to improve risk stratification. The goal of this study is to develop and test deep learning algorithms from diagnostic H&E images of RMS tumors which can aid in the diagnosis, mutation prediction and risk stratification for RMS patients. Methods: De-identified RMS patient samples were collected from tissue banking studies from Children’s Oncology Group (n=275), University Hospital Zurich (n=250) and Memorial Sloan Kettering (n=10). H&E stains on whole slides or TMAs were digitally scanned and used for analysis. Clinical information including clinical risk group, event-free survival, and genomic findings were used as available for training and testing. Convolutional neural networks (CNN) using EfficientNet were trained using K-fold cross validation to classify tumor histology, mutation probability, and risk stratification and tested against randomly selected samples or independent datasets when available. Results: The developed AI algorithm was able to classify tissue as ARMS (FP-RMS), ERMS (FN-RMS), stroma and necrosis with an average weighted intersect-over-union of 0.74 when compared to an expert pathologist annotation. A second deep learning algorithm developed specifically for distinguishing FP-RMS from FN-RMS displayed excellent sensitivity (FP-RMS=0.88) and specificity (FP-RMS=0.86) when tested against an independent RMS TMA dataset. Algorithms were also trained to predict mutations in MYOD1, RAS pathway genes, and TP53 and displayed good performance with ROC values of 0.96, 0.68, and 0.64, respectively. Lastly, we developed an algorithm to provide a Cox proportional hazard prediction based on H&E images. The resulting algorithm was capable of predicting EFS with similar accuracy as current clinical risk group assessment with improved ability to distinguish intermediate and high risk patients. Conclusions: Deep learning with convolutional neural networks provides pathologist-independent classification of RMS patients from simple H&E images. These AI algorithms can provide probabilities of prognostically relevant genetic alterations and survival which will ultimately contribute to better risk stratification of RMS patients. Citation Format: David Milewski, Hyun Jung, G. Thomas Brown, Yanling Liu, Jack Collins, Marc Ladanyi, Erin Rudzinski, Javed Khan. Predicting survival of rhabdomyosarcoma patients based on deep learning of H&E images [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 466.