Background/Aim: Current pathologic assessment of colorectal cancer (CRC) does not account for tumor heterogeneity, yet biologically important signals are embedded in morphologic differences that can drive patient prognosis. We utilized a deep learning algorithm (QuantCRC) to enhance patient risk stratification within DNA mismatch repair (MMR) groups [deficient (d) vs proficient (p) MMR]. QuantCRC quantifies 15 distinct tumor morphological features that were previously validated against interpretations by GI pathologists.
Materials and Methods: The training cohort consisted of 402 patients with resected stage III colon adenocarcinomas (191 d-MMR; 189 p-MMR) from a phase III trial of FOLFOX + cetuximab as adjuvant chemotherapy (N0147, Alliance) where study arms were pooled. The validation cohort consisted of 1275 stage III colon cancer patients (enrolled in the Colon Cancer Family Registry (CCFR) and at three academic sites) treated with fluoropyrimidine-based adjuvant chemotherapy (176 d-MMR; 1094 p-MMR). A quantitative segmentation algorithm (QuantCRC) recognizing 15 tumor morphological features was applied to H&E-stained whole slides that were digitalized (Figure 1). Association of these features with clinicopathologic variables, molecular alterations (KRAS, BRAF^V600E, MMR), and time-to-recurrence (TTR) was determined. Multivariable Cox proportional hazard models were developed to predict TTR within each MMR group.
Results: We found significant differences for quantitative morphological features by MMR status. Cancers with p-MMR had more immature desmoplastic stroma (p<0.001). Cancers with d-MMR had more inflammatory stroma (p<0.001 ) and higher epithelial TILs (p<0.001) in addition to increased high grade histology, mucin, and signet ring cells (all p <0.001). Neither desmoplastic nor inflammatory stromal subtypes differed by BRAF^V600E or KRAS status. In p-MMR tumors, multivariable analysis identified the tumor-stroma ratio as the strongest morphologic feature associated with TTR [dichotomous: HR_adj 2.02; 95% CI, 1.14-3.57; P=0.018], with 3-year recurrence rate of 40.2% vs 20.4% (Q1 vs Q2-4). Among d-MMR tumors, the extent of inflammatory stroma [continuous: HR_adj 0.98; 95% CI, 0.96-0.99; p=0.028; 3-year recurrence of 13.3% vs 33.4%; Q4 vs Q1] and N stage were the most robust prognostic features. The association of these features with clinical outcome was confirmed in the external validation cohort (Figure 2).
Conclusions: A deep learning algorithm (QuantCRC) applied to routine colon cancer sections enhances interpretive accuracy by identifying distinct morphological features driving patient prognosis within MMR groups. Our independently validated results can enable patient risk stratification for recurrence.

QuantCRC detects tumor morphological features in four layers. Abbreviations: TB/PDC, tumor budding/poorly differentiated cluster; TIL, tumor infiltrating lymphocytes.

AI-derived morphological features in stage III colon cancers (NCCTG N0147 trial) that were found to be most strongly associated with patient time-to-recurrence (TTR). Data are shown in Kaplan-Meier plots for tumor-stroma ratio among p-MMR tumors in A) training cohort B) validation cohort and % inflammatory stroma among d-MMR tumors in C) training cohort and D) validation cohort. Tumor stroma ratio is shown by level (level 1 < 0.65; level 2 > 0.65). Inflammatory stroma level includes level 1 (<7.38), level 2 (>7.38 and <15.22), level 3 (>15.22 and <36.75), level 4 (>36.75).

Kandavel Shanmugam

Appearances