Automatic Kidney Volume Estimation System Using Transfer Learning Techniques

Abstract

Deep learning technology is widely used in medicine. The automation of medical image classification and segmentation is essential and inevitable. This study proposes a transfer learning–based kidney segmentation model with an encoder–decoder architecture. Transfer learning was introduced through the utilization of the parameters from other organ segmentation models as the initial input parameters. The results indicated that the transfer learning–based method outperforms the single-organ segmentation model. Experiments with different encoders, such as ResNet-50 and VGG-16, were implemented under the same Unet structure. The proposed method using transfer learning under the ResNet-50 encoder achieved the best Dice score of 0.9689. The proposed model’s use of two public data sets from online competitions means that it requires fewer computing resources. The difference in Dice scores between our model and 3D Unet (Isensee) was less than 1%. The average difference between the estimated kidney volume and the ground truth was only 1.4%, reflecting a seven times higher accuracy than that of conventional kidney volume estimation in clinical medicine.