A CT-based interpretable deep learning signature for predicting PD-L1 expression in bladder cancer: a two-center study.

Abstract

Background: To construct and assess a deep learning (DL) signature that employs computed tomography imaging to predict the expression status of programmed cell death ligand 1 in patients with bladder cancer (BCa).

Methods: This retrospective study included 190 patients from two hospitals who underwent surgical removal of BCa (training set/external validation set, 127/63). We used convolutional neural network and radiomics machine learning technology to generate prediction models. We then compared the performance of the DL signature with the radiomics machine learning signature and selected the optimal signature to build a nomogram with the clinical model. Finally, the internal forecasting process of the DL signature was explained using Shapley additive explanation technology.

Results: On the external validation set, the DL signature had an area under the curve of 0.857 (95% confidence interval: 0.745-0.932), and demonstrated superior prediction performance in comparison with the other models. SHAP expression images revealed that the prediction of PD-L1 expression status is mainly influenced by the tumor edge region, particularly the area close to the bladder wall.

Conclusions: The DL signature performed well in comparison with other models and proved to be a valuable, dependable, and interpretable tool for predicting programmed cell death ligand 1 expression status in patients with BCa.