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

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Bladder cancer (BCa) management requires accurate prediction of programmed cell death ligand 1 (PD-L1) expression.
• Current methods for PD-L1 assessment can be invasive or time-consuming.

Purpose of the Study:

• To develop and validate a deep learning (DL) signature for predicting PD-L1 expression status in BCa patients using computed tomography (CT) imaging.
• To compare the performance of the DL signature against radiomics machine learning models.

Main Methods:

• Retrospective analysis of 190 BCa patients from two hospitals.
• Utilized convolutional neural networks and radiomics for model development.
• External validation and comparison with radiomics and clinical models; DL interpretability via Shapley additive explanation (SHAP).

Main Results:

• The DL signature achieved an area under the curve (AUC) of 0.857 on the external validation set, outperforming other models.
• SHAP analysis indicated that the tumor edge region, especially near the bladder wall, significantly influences PD-L1 prediction.
• The DL signature demonstrated superior predictive performance for PD-L1 expression status.

Conclusions:

• The developed DL signature is a valuable, dependable, and interpretable tool for predicting PD-L1 expression in BCa.
• This CT-based DL approach offers a non-invasive method for assessing PD-L1 status.
• The findings support the clinical utility of DL in precision oncology for bladder cancer.