PD-L1 expression assessment in Angiosarcoma improves with artificial intelligence support.

Tumoral PD-L1 expression is assessed to weigh immunotherapy options in the treatment of various types of cancer. To determine PD-L1 expression, each tumor cell needs to be assessed to calculate the percentage of PD-L1 positive tumor cells, called tumor proportion score (TPS). Pathologists cannot evaluate each cell individually due to time constraints and thus need to approximate TPS, which has been shown to result in low concordance rates. Decision quality could be improved by an AI-based TPS prediction tool which serves as a "second opinion". Establishing such a tool requires a certain amount of training data, which manifests a bottleneck for rare cancer types such as Angiosarcoma. To address this challenge, we developed and open sourced a pipeline that leverages pre-trained and generalist models to achieve strong TPS prediction performance on limited data. Pathologists were asked to reassess patients for which their TPS strongly disagreed with the AI's prediction. In many of these cases, pathologists updated their TPS score, improving their assessment, thus demonstrating the technical feasibility and practical value of AI-based TPS scoring assistance for rare cancers.