A rule-based approach as an alternative to end-to-end TIL prediction in gastroesophageal adenocarcinoma

David R Wessels, Ylva A. Weeda, Onno J. de Boer, Hoel Kervadec, Sybren L. Meijer, Erik J Bekkers

Keywords: Pathology, Computer Vision, Rule-Based

Abstract: Tumour-Infiltrating Lymphocytes (TILs) are a critical biomarker for predicting immunotherapy response in gastroesophageal adenocarcinoma (GEA). However, manual quantification of TILs on Whole Slide Images (WSIs) is labor-intensive and suffers from high inter-observer variability. While Deep Learning offers a solution for consistent assessment, current approaches typically fall into two distinct categories: interpretable, multi-step rule-based pipelines that mimic clinical workflows, or end-to-end Multiple Instance Learning (MIL) models that operate as "black boxes." These paradigms are rarely evaluated side-by-side. In this study, we propose a novel rule-based TIL quantification pipeline—comprising tissue segmentation and cell detection steps—and benchmark it against a standard end-to-end MIL approach. We utilize the AUMC-SELECT-AI dataset, a multi-center cohort combining TCGA and Amsterdam UMC data, to train our component models using dense ROI annotations. We evaluate both approaches on a held-out test set of 100 WSIs with pathologist-derived TIL scores. By comparing these methods, we assess the trade-offs between the granular interpretability of rule-based systems and the data-driven flexibility of end-to-end learning, providing insight into their clinical applicability for GEA biomarker assessment.

Primary Subject Area: Detection and Diagnosis

Secondary Subject Area: Application: Histopathology

Registration Requirement: Yes

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 296