Abstract 3514: A deep learning model for detection and characterization of tertiary lymphoid structures in H&E-stained images from pancreatic ductal adenocarcinoma

Abstract

Abstract Tertiary lymphoid structures (TLSs) are ectopic aggregates of a number of immune cells in nonlymphoid tissues under chronically inflamed environments and cancer. Emergence evidence suggested that TLSs were found in diverse cancers and TLSs have been referred as an independently predictive biomarker for immunotherapy response, especially immune checkpoint inhibitors. Moreover, the density and mature status of TLSs often positively associated with the favorable outcomes in cancers. Pancreatic ductal adenocarcinoma (PDAC) is one of the leading causes of cancer-related mortality with low survival rate in the whole world. Here, we developed a multi-resolution machine learning model for detection of TLSs from hema-toxylin and eosin (H&E) stained digital pathology slides as a potential clinical biomarker in PDAC. We selected 120 H&E slides (patients) with a total of 1,896 TLSs manually annotated as immature or mature TLSs by two pathologists. We also used immunohistochemistry (IHC) for CD20 and CD23 markers as a reference standard for H&E-based manual annotations. Our model mainly included detection and segmentation of individual TLSs. First, we used a transformer convolutional neural network to build a classification task to detect the presence of immature or mature TLSs. Then the instance segmentation model was applied to quantify the number of lymphocytes per unit square to validate the presence and maturation of TLSs. F1-score, calculated by the harmonic mean of precision (positive predictive value) and recall (sensitivity), was finally used to quantitatively evaluate the performance of model for each H&E slide. In validation dataset with 250 H&E slides (patients), we selected 50 slides to calculate F1-score, suggesting that our machine learning model achieved high F1-scores for detection and classification of immature TLSs (mean = 0.87, range from 0.79 to 0.95), and mature TLSs (mean = 0.92, range from 0.86 to 0.97). We further stratified 370 PDAC patients into three groups with mature TLSs (23%), immature TLSs (16%), and no TLSs (61%), respectively. Combined the clinical information, survival analysis revealed that patients with mature TLS had better survival outcomes than those without TLS (HR = 0.63; 95% CI, 0.44 - 0.90; P = 0.010). We developed a machine learning model that can accurately detect and classify TLSs in PDAC, and also demonstrated the prognosis value of predictive TLSs in H&E slides by automatedly machine learning model. These data highlight the promise of machine learning model for automated identification and quantification of the TLS on H&E slides in PDAC pathology. Citation Format: Chaoxian Zhao, Jidong Jia, Mingxi Lv, Jiayi Feng, Yijun Wang, Yingbin Liu. A deep learning model for detection and characterization of tertiary lymphoid structures in H&E-stained images from pancreatic ductal adenocarcinoma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 3514.