A method for predicting the outcome of PD1/PD-L1 inhibitors in non-small cell lung cancer

Objective: To propose a method for predicting immunotherapy outcome in non-small cell lung cancer based on computed tomography images before and after treatment. Methods: An improved U-net model incorporating Efficient Channel Attention was used to segment lesions. Radiomic features of lesions were extracted using PyRadiomics package and combined with biological indicators. Feature selection and dimensionality reduction were performed using linear discriminant analysis and Pearson correlation algorithms. A support vector machine was used to establish the predictive model. Results: The proposed segmentation model achieved a Dice coefficient of 90.09%, a positive predictive value of 89.23%, and an intersection over union of 82.15%, outperforming mainstream segmentation models. The proposed predictive model achieved an area under the curve of 85.05%, accuracy of 77.59%, specificity of 81.68% and sensitivity of 73.52%, all superior to models based solely on single-time computed tomography images or lacking biological features. Conclusion: The proposed method provides an effective approach for predicting the efficacy of immunotherapy in non-small cell lung cancer patients and offers a  valuable tool to support clinical decision-making.