Deep learning predicts microsatellite instability status in colorectal carcinoma in an ethnically heterogeneous population in South Africa

Background Deep learning (DL) models are effective pre-screening tools for detecting mismatch repair deficiency (dMMR) in colorectal carcinoma (CRC). These models have been trained and validated on large cohorts from the Northern Hemisphere, without representation of African samples. We sought to determine the performance of a DL model in an ethnically heterogeneous cohort of patients from South Africa. Methods Our cohort comprised 197 CRC resection specimens, with scanned whole slide images tessellated and inputted into a transformer-based DL model trained on large international cohorts. Model performance was evaluated using area under the receiver operating characteristic curve (AUROC), sensitivity and specificity. The maximal Youden’s J index was calculated to determine the optimal cut-off threshold for the model prediction score. Results Our model yielded an AUROC of 0.91 (±0.05). Using a prediction score threshold of 0.620 produced an overall sensitivity of 85.7% (95% CI 73.3% to 92.9%) and a specificity of 82.4% (95% CI 75.5% to 87.7%). The false negative cases were predominantly left-sided (71.4%) and did not show the typical dMMR/microsatellite instability-high histological phenotype. Sensitivity was lower (50%–75%) in cases showing isolated PMS2 or MSH6 loss of staining. Calibrating the classification threshold to 0.470, the sensitivity was optimised to 95.6% (95% CI 86.3% to 98.9%) with a specificity of 69.6% (95% CI 61.8% to 76.4%). This would have resulted in excluding 103 cases (52.3%) from downstream immunohistochemical (IHC) or molecular testing. Conclusions Following appropriate region-specific calibration, we have shown that this model could be employed to accurately prescreen for dMMR in CRC, thereby reducing the burden of downstream IHC and molecular testing in a resource-limited setting.