Deep Learning based Morphometric Analysis of BAL Cells for Clinical Outcome Assessment in Lung Transplant Recipients

Background: Chronic lung allograft dysfunction (CLAD) remains a major obstacle to long-term survival after lung transplantation. Routine clinical surveillance of lung allografts using bronchoalveolar lavage (BAL) cytology primarily relies on differential cell counts, potentially overlooking valuable morphological information. This study evaluates whether morphometric features derived from BAL cells can serve as prognostic markers for CLAD and graft survival. Methods: BAL cytospin images from 60 lung transplant recipients at the LMU Klinikum were analyzed. A U-Net model was established for nuclei segmentation and applied to outputs from a prior Mask R-CNN cell classification and segmentation pipeline. Morphometric features, including cell and nuclear area, perimeter, compactness, solidity, aspect ratio, and nucleus-to-cell ratio, were extracted. Iterative, visually guided thresholding was used to filter out artifacts from macrophages. Morphometric data were merged with clinical records and analyzed for associations with CLAD onset and graft survival using Kaplan-Meier, log-rank, and Cox models. Results: The nuclei segmentation model performed robustly across cell types (median IoU of 0.89 for macrophages). Morphometric features, such as solidity, compactness, aspect ratio and nucleus-to-cell ratio, effectively distinguished artifacts from valid predictions. Larger median macrophage area and perimeter were significantly associated with improved CLAD-free survival and reduced risk of developing CLAD, while greater median perimeter also correlated with better graft survival. Other morphometrics showed no significant association with outcomes. Conclusion: This study demonstrates the feasibility of morphometric analysis of BAL cytospin slides for low-cost risk assessment of CLAD in lung transplant recipients. Macrophage size metrics, particularly cell area and perimeter, emerged as potential biomarkers for CLAD risk and graft survival. Although promising, these findings require validation in larger multicenter cohorts and integration into digital cytology workflows before clinical translation. Publication History Article published online: 04 March 2026 © 2026. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany