pISSN 3022-6783
eISSN 3022-7712

View

Article View

Korean J Transplant 2023; 37(Suppl 1): S8-S8

Published online November 15, 2023

https://doi.org/10.4285/ATW2023.F-5666

© The Korean Society for Transplantation

Development and validation of ultrasonography-based deep learning models for prediction of rejection in kidney transplant patient

Yoo Jin Lee, Yang Wook Kim, Chang Min Heo, Sihyung Park, Bong Soo Park

Department of Nephrology, Inje University Haeundae Paik Hospital, Busan, Korea

Correspondence to: Yang Wook Kim
E-mail: kyw8625@chol.com

Abstract

Background: About 15%–20% of kidney transplant recipients experience rejection. Acute rejection is the most important cause of transplant failure after kidney transplantation, and differential diagnosis of acute transplant dysfunction remains a difficult clinical challenge. Kidney biopsy for diagnosis of rejection has the risks such as bleeding and infection. We predicted rejection using a deep learning-based model using ultrasonography images of transplanted kidneys.
Methods: Among patients who underwent kidney transplantation between March 2010 and September 2022, patients who underwent a biopsy after transplantation were classified into cellular rejection group (G1), antibody-mediated rejection group (G2), and nonrejection group (G3). Ultrasonography of transplanted kidneys performed before biopsy was used. Using deep learning, 356 images of G1, 544 images of G2, and 790 images of G3 were analyzed. First, it was analyzed whether it could be predicted by distinguishing G1, G2, and G3, and second, whether it could be predicted by distinguishing the group with rejection and the group without rejection.
Results: Using the ResNet model pretrained with the ImageNet dataset, the model training result, which analyzed whether predictions could be made by dividing Gl and G2 and the group without rejection, showed an accuracy of 88.76%. In the same way, the model learning result, which analyzed whether it could be predicted by dividing the group with and without rejection, showed an accuracy of 93.33%. However, when checked with GradCAM, activation results were shown in the upper part outside the kidney area and in the letter area. It is thought that it is necessary to make a decision on data refinement and model advancement based on GradCAM and performance figures.
Conclusions: We confirmed the possibility of predicting the biopsy results through ultrasound images of transplanted kidneys using deep learning. It is expected that additional studies conducted through enhancement through more sophisticated images will be needed.