Abstract LB117: AI for decoding spatial organization and molecular convergence in angiosarcoma

Abstract Angiosarcomas are a heterogeneous group of aggressive sarcomas that form malignant blood vessels. They are rare in humans, with an annual incidence of 500-1,000 individuals in the United States. In contrast, these malignancies, also called hemangiosarcomas, frequently occur in domestic dogs, with an estimated 50,000 cases annually. In both species, angiosarcomas create malignant endothelium characterized by irregular vascular channels containing red and white blood cells. Precise identification of malignant endothelial cells is challenging due to the complex admixture of tumor cells with adjacent stromal components, their similar mesenchymal morphology, and severe intratumoral thrombosis. The spatial orientation of each tissue component is poorly characterized, and its pathogenic significance is unknown.The aim of this study is to decode spatial histology and facilitate the functional estimation of angiosarcomas by leveraging artificial intelligence (AI) techniques. Whole slide images (WSIs) were generated from visceral angiosarcomas (n=70) and benign diseases (n=30), including splenic hematoma and lymphoid hyperplasia, obtained from client-owned dogs. Additionally, independent datasets were created by manually capturing representative histological images from the tumors (n=4,316) and benign lesions (n=3,536), providing image-level annotation for weakly supervised learning.We first developed convolutional neural network (CNN) models using ResNet-50 and AlexNet to classify weakly-annotated histological images of angiosarcomas from benign images, achieving Area Under the Receiver Operating Characteristic Curve values of 0.999 and 0.998, respectively. An explainable AI tool, Grad-CAM, highlighted image components representing endothelial malignant tumor components in a pathologically relevant manner. Then, WSI datasets were patched, clustered into distinct image groups, and transformed into digital heatmaps using Clustering-constrained Attention Multiple Instance Learning architecture. The heatmaps illustrated the spatial organization of multiple tissue components, portraying irregular and deforming malignant endothelial structures along with the microenvironmental components. Our prototypic AI models captured a comprehensive picture of spatial organization, which was transformed into digital metrics. This technique identified five major spatial characteristics that recapitulate distinct histological features depicted on the heatmaps, and the decoded matrices pinpoint spatial locations. Ongoing work is developing AI models that enable functional estimation of each component by incorporating genomic and molecular data. Citation Format: Donghee Lee, Rohith Sai Reddy Kondreddy, Hyunji Jo, Jong Hyuk Kim. AI for decoding spatial organization and molecular convergence in angiosarcoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 2 (Late-Breaking, Clinical Trial, and Invited Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_2):Abstract nr LB117.