DEEP LEARNING TO PREDICT DEGREE OF INTERSTITIAL FIBROSIS AND TUBULAR ATROPHY FROM KIDNEY ULTRASOUND IMAGES - AN ARTIFICIAL INTELLIGENCE APPROACH

Athavale, A. M.; Hart, P. D.; Itteera, M.; Patel, T.; Cymbaluk, D. J.; Alabka, A.; Dunae, G.; Arruda, J.; Singh, A.; Rosenberg, A.; Kulkarni, H.

Background: Interstitial fibrosis and tubular atrophy (IFTA) is a strong predictor of decline in kidney function. Non-invasive test to assess IFTA is not available. Methods: We trained, validated and tested a deep learning (DL) system to classify IFTA grade from 6,135 ultrasound images obtained from 352 patients who underwent kidney biopsy. Of 6,135 ultrasound images, 5,523 were used for training (n = 5,122) and validation (n = 401) and 612 to test the accuracy of the DL system. IFTA grade scored by nephropathologist on trichrome stained kidney biopsy slide was used as reference standard. Results: There were 159 patients (2,701 ultrasound images), 74 patients (1,239 ultrasound images), 41 patients (701 ultrasound images) and 78 patients (1,494 ultrasound images) with IFTA grades 1, 2, 3 and 4, respectively. The deep-learning classification system used masked images based on a 91% accurate kidney segmentation routine. The performance matrices for the deep learning classifier algorithm in the validation set showed excellent precision (90%), recall (76%), accuracy (84%) and F1-score (80%). In the independent test set also, performance matrices showed excellent precision (90%), recall (80%), accuracy (87%) and F1-score of (84%). Accuracy was highest for IFTA grade 1 (98%) and IFTA grade 4 (82%). Conclusion: A DL system can accurately predict IFTA from kidney ultrasound image.