Development and Validation of a Computed Tomography–Based Convolutional Neural Network Model for Assisting Donor Lung Acquisition in Lung Transplantation

Lung transplantation remains the only effective treatment for end-stage pulmonary disease. However, transplantation rates are significantly constrained by the ongoing shortage of donor lungs and the reliance on subjective assessments of donor lung quality. Although computed tomography (CT) is essential for donor lung evaluation, there is currently limited evidence supporting the use of automated prediction algorithms to determine donor lung suitability. This retrospective study included potential donors evaluated between May 2017 and July 2023 at the First Affiliated Hospital of Guangzhou Medical University, where donor lung CT images and utilization outcomes were collected. A convolutional neural network (CNN) was developed and evaluated using patient-level stratified splitting, with 70% of donors assigned to the training set and 30% assigned to the validation set. Model performance was summarized by AUC and 95% confidence intervals estimated via 1000 bootstrap resamples of the validation set. A total of 111 donor lungs underwent CT-based evaluation for transplantation; of these, 21 (18.9%) were ultimately transplanted and 90 (81.1%) were not. The CNN model achieved an AUC of 0.839 (95% CI 0.658–0.968) and an accuracy of 0.795 (95% CI 0.767–0.900) in predicting donor lung utilization. This exploratory single-center study suggests that deep learning applied to donor lung CT imaging may enable prediction of donor lung utilization with promising preliminary performance. Given the small and imbalanced cohort, these findings should be interpreted cautiously and require validation in larger, prospective, multicenter studies.