Machine Learning-Based Ultrasound Radiomics for Evaluating the Function of Transplanted Kidneys

The aim of the study described here was to investigate the value of different machine learning models based on the clinical and radiomic features of 2-D ultrasound images to evaluate post-transplant renal function (pTRF). We included 233 patients who underwent ultrasound examination after renal transplantation and divided them into the normal pTRF group (group 1) and the abnormal pTRF group (group 2) based on their estimated glomerular filtration rates. The patients with abnormal pTRF were further subdivided into the non-severe renal function impairment group (group 2A) and the severe impairment group (group 2B). The radiomic features were extracted from the 2-D ultrasound images of each case. The clinical and ultrasound image features as well as radiomic features from the training set were selected, and then five machine learning algorithms were used to construct models for evaluating pTRF. Receiver operating characteristic curves were used to evaluate the discriminatory ability of each model. A total of 19 radiomic features and one clinical feature (age) were retained for discriminating group 1 from group 2. The area under the receiver operating characteristic curve (AUC) values of the models ranged from 0.788 to 0.839 in the test set, and no significant differences were found between the models (all p values >0.05). A total of 17 radiomic features and 1 ultrasound image feature (thickness) were retained for discriminating group 2A from group 2B. The AUC values of the models ranged from 0.689 to 0.772, and no significant differences were found between the models (all p values >0.05). Machine learning models based on clinical and ultrasound image features, as well as radiomics features, from 2-D ultrasound images can be used to evaluate pTRF.