Revolutionizing outcome prediction for immunotherapy with vision transformer-based pathomics signature in esophageal squamous cell cancer.

Abstract

e16049 Background: Immune checkpoint inhibitors (ICIs) has notably enhanced the survival prospects for patients with esophageal squamous cell carcinoma (ESCC), however the clinical benefit was limited to only a small portion of patients. This study aimed to perform a deep learning signature based on H&E-stained pathological specimens for the purpose of identifying effective markers that can accurately predict the clinical benefit of ICIs among patients with ESCC. Methods: ESCC patients receiving PD-1 inhibitors from Shandong Cancer Hospital were included. Whole slide images (WSIs) of H&E-stained histological specimens of included patients were collected, and subsequently randomly allocated into two sets: a training set comprising 70% of the samples, and a validation set consisting of the remaining 30%. The labels of images were defined by the progression-free survival (PFS) with the interval of 4 months. The pretrained ViT model was used for patch-level model training, and all patches were projected into probabilities after linear classifier. Then the most predictive patches were passed to RNN for final patient-level prediction to construct ESCC-pathomics signature (ESCC-PS). Accuracy rate and survival analysis were performed to evaluate the performance of ViT-RNN survival model in validation cohort. Results: 163 ESCC patients receiving PD-1 inhibitors were included for model training. There were 486,188 patches of 1024*1024 pixels from 324 WSI images of H&E-stained histological specimens after image pre-processing. The training cohort consisted of 120 patients with a total of 227 images, while the validation cohort comprised 43 patients and featured 97 images. Notably, both groups exhibited balanced baseline characteristics between them. The ESCC-PS achieved an accuracy of 84.5% in the validation cohort, and could distinguish patients into three risk groups with the median PFS of 2.7, 4.8 and 16.7 months (P < 0.001). The multivariate cox analysis revealed ESCC-PS could act as an independent predictor of survival from PD-1 inhibitors (P < 0.001). A combined signature incorporating ESCC-PS and expression of PD-L1 shows significantly improved accuracy in outcome prediction of PD-1 inhibitors compared to ESCC-PS and PD-L1 anlone, with the area under curve value of 0.952, 0.924, 0.648 for 6-month PFS and C-index of 0.816, 0.806, 0.66, respectively. Conclusions: The outcome supervised pathomics signature based on deep learning has the potential to enable superior prognostic stratification of ESCC patients receiving PD-1 inhibitors, which convert the images pixels to an effective and labour-saving tool to optimize clinical management of ESCC patients.