Automated tertiary lymphoid structure detection and prognostic value in colorectal liver metastases

Background: Tertiary lymphoid structures (TLS) are organized lymphoid cell aggregates that can form within solid cancers and are generally associated with more favorable prognosis.1 TLS significance in colorectal liver metastases (CRLM) remains unclear,2 while reproducible and scalable tools for TLS characterization on routine histopathology are lacking. We developed an automated machine-learning pipeline to detect and classify TLS from hematoxylin and eosin (H&E) WSI and evaluated association with overall survival (OS) in patients who underwent CRLM resection with curative intent.

Methods: As ground truth, a TCGA gastric adenocarcinoma cohort (n = 76) with a dataset of 235 TLS expert-annotated in three morphologic classes on H&E images was used to train and tune the pipeline, then adapted to our internal cohort of 1286 CRLM resected in 314 patients. After stain-normalization of H&E slides, a multiresolution latent hypergraph learning framework was applied, incorporating a diffusion-based model at two magnifications to capture tissue-level context (10x) and cell-level graphs (40x). TLS were categorized into three morphologic classes: TLS1, lymphoid aggregates; TLS2, immature; TLS3, with germinal center (figure 1). Each patient received a score computed as a weighted sum of the three TLS subtypes, and assigned to groups: TLS score = 0 (Absent), TLS score > 0 to ≤ Q3 (Intermediate), and TLS score > Q3 (High), where Q3 represented the 75th percentile of TLS scores across the cohort. Association with OS was assessed using KM analysis and the log-rank test.

Results: The pipeline developed with the TCGA dataset correctly assigned TLS to morphological classes with an accuracy of 91% , and TLS scoring stratified patients by overall survival. Out of a total of 4855 TLS detected in CRLM, 3056 (62.9%), 1328 (27.4%), and 471 (9.7%) were TLS1, TLS2, and TLS3, respectively. No TLS were found in 215 CRLM, representing 16.7% patients. The median OS of patients with Absent TLS, Intermediate, and High scores were 30, 45, and 65 months, respectively (p < .001) (figure 1F). Pairwise comparisons confirmed significant differences across all strata (0 vs. intermediate, p < .001; 0 vs. high, p < .001; intermediate vs. high, p = .03). A greater proportion of patients who received pre-operative chemotherapy were found in TLS High compared to TLS Absent and Intermediate groups (p = 0.01).

Conclusions: We developed a graph-based computational pathology pipeline for automated TLS scoring from H&E-stained CRLM. Validation of its prognostic value in independent cohorts could ultimately help clinical decision-making for adjuvant therapy.