Noninvasive monitoring of allograft rejection in a rat lung transplant model: Application of machine learning-based 18F-fluorodeoxyglucose positron emission tomography radiomics

Standardized uptake values (SUVs) derived from 18F-fluorodeoxyglucose (18F-FDG) positron emission tomography (PET) are valuable but insufficient for detecting lung allograft rejection (AR). Using a rat lung transplantation (LTx) model, we investigated correlations of AR with the SUVmax and PET-derived radiomics and further evaluated the performance of machine learning (ML)-based radiomics for monitoring AR.LTx was performed on 4 groups of rats: isograft, allograft-cyclosporinecontinuous (CsAcont), allograft-CsAdelayed, and allograft-CsA1week. Each rat underwent 18F-FDG PET at week 3 or 6. The SUVmax and radiomic features were extracted from the PET images. Least absolute shrinkage and selection operator regression was used to construct a radiomics score (Rad-score). Ten modeling algorithms with 7 feature selection methods were performed to develop 70 radiomics models (49 ML models and 21 logistic regression models) for monitoring AR, validated using the bootstrap method.In total, 837 radiomic features were extracted from each PET image. The SUVmax and Rad-score showed significant positive correlations with histopathology (p < .05). The area under the curve (AUC) of SUVmax for detecting AR was 0.783. The median AUC of ML models was 0.921, which was superior to that of logistic regression models (median AUC, 0.721). The optimal ML model using a random forest modeling algorithm with random forest feature selection method exhibited the highest AUC of 0.982 (95% confidence interval, 0.875-1.000) in all models.SUVmax provided a good correlation with AR, but ML-based PET radiomics further strengthened the power of 18F-FDG PET functional imaging for monitoring AR in LTx.