Developing a deep learning algorithm to support decision making for the classification of microsatellite instability (MSI) in cancer, using artificial intelligence(AI).

11152 Background: Digital pathology (DP) is an image-based environment, which allows the acquisition, and interpretation of pathological information from a digitized slide. The objective of this work was to build an AI model for predicting MSI in endometrial, colorectal, and gastric carcinomas, to simplify laboratory processes, in a developing country. Nowadays, it is crucial to recognize patients who will respond to treatments such as immunotherapy. Methods: The computational tool utilized was CLAM (clustering-constrained-attention multiple instance learning). This architecture developed by Lu et al. is a deep learning method that utilizes weak supervision and attention based learning, to identify the most relevant regions for tumor classification, in whole slide images (WSI). CLAM was designed to excel in tumor subtyping tasks which are discernible to human visual perception. We are challenging this architecture by employing it in the detection of a molecular phenomenon. For training and internal evaluation, images from TCGA and CPTAC were obtained for endometrial cancer (EC). For colorectal (CCR) and gastric cancer (GC), only images from TCGA. Archived histological samples with MSI/MMR status from a donor lab , were used for external validation. Results: Tests conducted on an internal cohort of 58 samples for CCR and GC revealed a sensitivity of 76.9%, a specificity of 78.1%, and an AUC-ROC of 76.3%. In the case of EC, where 38 cases were evaluated, the sensitivity was 100%, specificity was 68%, and AUC-ROC reached 92%. Notably, regarding the externalcohort consisting of 109 samples, both models demonstrated a high sensitivity of up to 90% in recognizing positive cases. Due to the models’ specificity (approximately 45%), it was not possible to accurately predict MSS samples in all three types of tumors. Conclusions: Promising results were obtained for the first approach in detecting molecular biology events, such as MSI, using AI. CLAM is a useful platform for digital pathology, but new validations are needed to enhance our work and strengthen the model. It is hypothesized that CLAM may not be the best tool for AI in identifying molecular biology events, so other directions are being explored. For external cohorts, preanalytical changes play an important role in obtaining accurate results with scanners and digitalization. It is crucial to teach healthcare professionals about AI in order to ensure its availability, in every diagnostic laboratory.