Abstract 7388: Deep learning-implemented molecular signature identification in vascular malignancies

Abstract

Abstract Angiosarcoma is a rare, aggressive sarcoma that forms malignant endothelium. Naturally occurring angiosarcoma in domestic dogs, also called hemangiosarcoma, is relatively common and offer a useful translational model for the rare human cancer. Distinct molecular subtypes of angiosarcoma have been identified; however, current methods are cost-prohibitive and time-consuming, posing challenges in the development of clinical applications. In this study, we hypothesize that artificial intelligence enables reliable image-based diagnostics and molecular characterization specific to angiosarcoma. The aims of our work are to develop a deep learning framework to provide diagnostic support for suspicious malignant cases, to detect morphological features associated with oncogenic mutations, and to identify immune phenotypes in this malignancy. For proof-of-concept and feasibility study, we first utilized convolutional neural network models to identify malignant features in angiosarcomas. We curated 4,079 images of hematoxylin and eosin (H&E) stained slides from seventy visceral angiosarcomas and 3,536 images of thirty splenic hematomas as benign controls obtained from client-owned dogs. ResNet-50 and AlexNet were trained on the images, and area under the receiver operating characteristic curve (AUC) was determined for disease classification. ResNet-50 and AlexNet yielded AUCs of 0.999 and 0.998, respectively. Next, we calculated the gradient of the model’s prediction and applied it to the original image using Grad-CAM software. Grad-CAM results revealed that AlexNet highlighted histological regions representing malignant cellular components for decision-making, aligning with the pathological diagnostic context. In addition, we applied ResNet-50 to determine if the model could predict morphological features in angiosarcoma tissues representing the presence of PIK3CA mutation. We employed RNA-seq data to annotate PIK3CA mutation status on the matched histological images. Among 54 cases annotated with PIK3CA mutation status, 2,732 PIK3CA wild-type and 664 mutant tumor images were trained by ResNet-50. Our model demonstrated an AUC of 0.981 in the prediction of PIK3CA mutation. We further defined immunophenotypic groups using xCell algorithms on RNA-seq data, and immune scores were applied to histological images. ResNet-50 model resulted in an AUC of 0.953 to classify immune-high and immune-low angiosarcomas. RNA-seq datasets generated from 13 human angiosarcoma cases exist in our lab, along with the respective H&E slides. We also have immunohistochemical slides available for p53, phospho-p53 (Ser15 and Ser20), AKT, and phospho-AKT (Thr308) proteins conducted in both human and canine angiosarcoma tissues. Our ongoing work involves generating whole-slide images of H&E stains and immunostained slides to enhance the deep learning models for the molecular classification of angiosarcomas. Citation Format: Donghee Lee, Heon Heo, Joanne Kim, Jong Hyuk Kim. Deep learning-implemented molecular signature identification in vascular malignancies [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7388.