Machine-learning based BAL cytology of lung transplant patients

Introduction: Long-term survival after lung transplantation (LTX) is severely limited by the poorly understood chronic lung allograft dysfunction (CLAD). Previous studies identified bronchoalveolar lavage (BAL) neutrophilia and eosinophilia as risk factors for CLAD. Usually, BAL differential cytology is manually assessed, which is time-consuming and examiner dependent. We utilized a machine-learning based BAL cytology computer vision (BAL-ML) model to generate differential cytology of BAL cytospins of LTX patients. Methods: 575 Diff-Quik-stained BAL cytospins of LTX patients were imaged using a Zeiss AxioScan slide scanner. Manual cell segmentation and cell labeling for the ground truth generation were performed by two respiratory physicians independently. The labeled training data contained representative as well as granulocyte- and lymphocyte-enriched image-subsets. Results: Preliminary Mask-RCNN based models performed comparable to human interrater reliability in macrophage, lymphocyte and granulocyte evaluation. Challenges by imbalances in cell distributions were seen using representative training image data only. These were overcome by the addition of granulocyte- and lymphocyte-enriched training datasets. Conclusions: BAL differential cytology by BAL-ML models is feasible and accurate. Naturally occurring label imbalances in cytospin image data can be overcome by oversampling granulocyte and lymphocyte occurrences in the training data. BAL-ML models will provide deeper insights into associations between differential cytology and morphometric features of BAL cytology with CLAD development in lung transplant patients. In future, validated models could assist BAL cytology in clinical routine.