569P - Artificial intelligence for detection of mismatch repair deficiency in colon carcinomas (alliance)

Background

Artificial intelligence (AI) algorithms have been developed to distinguish MMR-D from MMR-P (proficient) tumors, yet they lack biological interpretability. To address this shortcoming, we used validated AI algorithms to detect and quantify well described morphological features in colon cancers from whole slide images.

Methods

Digital images of stage III colon carcinomas from participants in a phase III trial of FOLFOX-based adjuvant chemotherapy (NCCTG N0147; Alliance) were utilized. Available MMR-D (n=191) and randomly selected MMR-P (n=188) cancers were analyzed. QuantCRC quantifies the amount of stroma and stromal subtypes (immature, mature, inflammatory) and tumor/stroma ratio in addition to grade, mucin, signet ring cells, budding/poorly differentiated clusters (TB/PDC) and tumor-infiltrating lymphocytes (TILs). We also distinguished TILs in epithelium vs stromal using another AI algorithm (Lunit SCOPE IO). Area under the curve (AUC), i.e., discrimination threshold, was measured for each feature to predict MMR status.

Results

Among AI-derived morphological features, significant differences were found by MMR status whereby dMMR cancers had lower immature stroma but higher inflammatory stroma, higher tumor grade, and increased mucin, signet ring cells, tumor size, TB/PDCs, and TILs (all Kruskal-Wallis p<0.05). Univariately, AUC (95% CI) values for prediction of MMR status above 0.6 were as follows: TILs 0.72 (0.67-0.77), immature stroma 0.67 (0.62-0.73); inflammatory stroma 0.67 (0.62-0.73); tumor grade 0.65 (0.60-0.71); signet ring cells 0.65 (0.60-0.71); mucin 0.64 (0.59-0.70). Given that TILs in the overall tumor was most discriminant, we then utilized a TIL-specific AI algorithm measuring TILs in epithelium and stroma and found that epithelial TILs had a significantly higher AUC [ 0.79 (0.74-0.84) compared to stromal TILs 0.58 (0.52-0.63) p<0.0001].

Conclusions

Among morphological tumor features quantified by AI, tumor immature desmoplastic stroma and TILs, particularly epithelial TIL densities, were best able to distinguish MMR-D from MMR-P tumors. These findings have implications for prognosis and subtype-based interventions and can guide spatial molecular interrogation.

Clinical trial identification

NCT00079274.

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Supported by Institutional grant support. NIH U10CA180821, U10CA180882, U24CA196171, R01 CA210509 (to FAS). Study NCCTG N0147 received funds from Sanofi.

Disclosure

G. Park, Y. Lim: Financial Interests, Institutional, Member: Lunit. D. Yan, K. Shanmugam: Financial Interests, Institutional, Full or part-time Employment: Roche. C-Y. Ock: Financial Interests, Institutional, Member: Lunit. All other authors have declared no conflicts of interest.