Meejeong Kim, Philip Chikontwe, Heounjeong Go, Jae Hoon Jeong, Su-Jin Shin4, Sang Hyun Park*, Soo Jeong Nam*
Background: Microsatellite Instability (MSI) is a clinically significant subtype in colorectal cancer. Despite the promising performance of deep learning techniques in digital pathology for clinical diagnosis, the impact of clinicopathologic factors on the performance of these models has been largely overlooked.
Methodology: Using a total of 931 colorectal cancer Whole Slide Images (WSIs), we developed and verified a deep learning algorithm and analyzed the WSI-level MSI probability and clinicopathologic variables.
Results: In both internal and external cohorts, our deep learning model achieved an Area Under the Receiver Operating Curve (AUROC) of 0.901 and 0.908, respectively. The presence of a mucinous or a signet ring cell carcinoma component enhanced the model’s ability to predict MSI (HR=19.73, P=0.026). Conversely, tumors subjected to neoadjuvant chemoradiation therapy (HR=0.03, P=0.002) and those with metastasis (HR=0.01, P=0.016) demonstrated an increased probability of being associated with Microsatellite Stability (MSS).
Conclusion: To ensure the clinical applicability of the model, it is imperative to meticulously validate deep learning- based approaches for MSI prediction, accounting for diverse practical clinicopathologic backgrounds that may impact the model’s performance.