Deep learning–powered analysis of tumor-infiltrating lymphocytes (TILs) in colorectal cancer.

288 Background: TILs are a promising biomarker in various cancers, including colorectal cancer (CRC). Advances in deep learning enable objective and efficient TIL assessment. This study applies a novel deep learning model to quantify TIL density in a large cohort of CRC tumors, examining its relationship with clinicopathologic factors and survival. Methods: We applied a deep learning model (Deep-ICP) to H&E-stained images from CRC patients treated at Dana-Farber Cancer Center to quantify TIL density (immune cells/mm²). Kruskal-Wallis and Wilcoxon tests were used to compare categories. Overall survival (OS) was calculated from the date of diagnosis to the date of death or last follow-up. Cox proportional hazards models analyzed the association between TIL density tertiles (lower, middle, upper) and OS, adjusting for age, sex, stage, biopsy site, tumor mutational burden (TMB, mutations/megabase), and mismatch repair (MMR) status. Results: Of 1,477 tumor specimens tested, 1,039 (70%) were primary and 438 were metastatic. The most common sites of metastasis were liver (n=205), lung (n=86), and soft tissue (n=35). The median age at diagnosis was 56 (range 19–91) years and 706 were female (48%). Most tumors (1,214/1,300; 93%) were MMR proficient (MMRp); 86 were MMR deficient (MMRd). Median TIL density was 664 (interquartile range; 469–934) cells/mm² in primary tumors, 478 (358–685) cells/mm² in liver metastases, 415 (251–645) cells/mm2 in soft tissue (p<0.001). MMRd tumors had higher median TIL density than MMRp (881 vs. 620 cells/mm²; p<0.001). After adjusting for covariates, patients with highest TIL tertile had significantly longer OS than patients with lowest TIL tertile (median OS: 75.4 [95%CI 67.4-87] months; 100.5 [83.9-116] months and 99.1 [92.4-157] months for lower, middle and upper TIL tertiles respectively; Table). In the MMRp subset, higher TIL density was still predictive of improved OS (p adj=0.010). Conclusions: This study demonstrates the feasibility of deep learning powered TIL quantification in CRC, potentially informing differences in tumor microenvironment and prognostication. Future prospective studies are needed to validate the results and explore their therapeutic implications in personalized CRC management. Variable Hazard Ratio (95% CI) P-value TIL upper tertile (vs. lower) 0.70 (0.50-0.99) 0.04 TIL middle tertile (vs. lower) 1.04 (0.77-1.39) 0.80 Age (continuous) 1.02 (1.00-1.03) 0.008 Male (vs. female) 1.13 (0.86-1.47) 0.38 Stage II (vs. stage I) 4.8 (0.59-40) 0.14 Stage III (vs. stage I) 9.4 (1.20-73) 0.03 Stage IVA (vs. stage I) 23 (2.94-178) 0.003 Stage IVB (vs. stage I) 46 (5.6-369) 0.0003 Liver specimen (vs. primary) 1.39 (1.00-1.92) 0.050 Lung specimen (vs. primary) 0.78 (0.47-1.29) 0.33 Lymph node specimen (vs. primary) 0.77 (0.27-2.17) 0.62 Other metastasis (vs. primary) 1.23 (0.78-1.93) 0.37 TMB≥10 (vs. <10) 1.17 (0.82-1.68) 0.38 MMRd (vs. MMRp) 0.65 (0.29-1.46) 0.30