Evaluation of Transfer Learning with a U-Net Architectures for Kidney Segmentation

Kidney cancer emerges as one of the primary causes of mortality due to neoplasms on a global scale. Early detection and diagnosis of this disease often allow for more treatment options, contributing to the reduction of death rates. In this way, a correct delimitation of kidneys and renal tumor areas provides better analysis and diagnosis of suspicious lesions, contributing to treatment planning. This task is usually performed manually, making the process susceptible to fatigue (physical and visual) and distraction. Therefore, computational techniques, such as deep neural networks, are presented with great prominence as alternatives to improve segmentation precision and contribute to the early diagnosis of kidney cancer. In this work, we propose a methodology for kidney segmentation in computed tomography images by transfer learning to the U-Net network architecture. The KiTS19 dataset was used to evaluate the proposed methodology and obtained the best result for kidney segmentation of 96.0% of average Dice coefficient and average Jaccard index of 94.4%, using a pre-trained EfficentNet as an encoder for a U-Net.