Deep learning segmentation of kidneys with renal cell carcinoma.

Abstract

e16098 Background: Renal cell carcinoma (RCC) is the most common kidney malignancy and the eighth most common cancer overall, with more than 60,000 new cases and nearly 15,000 deaths in the United States each year. The majority of RCC is asymptomatic and detected incidentally on computed tomography (CT) scans performed for other indications. Multiple prior studies have highlighted the potential for a lesion to be missed or mischaracterized. Artificial intelligence (AI), the mimicking of human cognition by computers, has promise for improving radiological accuracy. Convolutional neural networks (CNNs), a type of AI, apply a feed-forward design to an artificial neural network. Multi-layered CNNs thus have application in biomedical imaging and radiology. In this study we describe an approach for automatic localization and segmentation of the kidneys on contrast enhanced CT scan. Methods: This IRB approved retrospective study included patients who were diagnosed with RCC and had an abdominal CT scan with IV contrast from 2011 to 2019. For each abdominal CT scan, the kidneys and any renal lesions were manually segmented by a board-certified abdominal radiologist on the coronal post-contrast phase of the CT scan. Two CNNs were developed and trained to segment the kidneys automatically. Initially, the CNN implemented a regression algorithm that predicted a 3D bounding box around each kidney. Subsequently, the second CNN used a 3D/3D U-Net algorithm that received the bounding cube as its input and outlined the kidney as an output. After training, the CNN algorithms were assessed by a Sørensen-Dice coefficient. Results: The first CNN that predicted a 3D bounding box around each kidney was trained on 127 abdominal CT exams. The second CNN that segmented the kidneys with RCC was trained on 259 bounding boxes that encompassed the kidneys. The Sørensen–Dice coefficient for segmentation of the kidneys with RCC was 0.92. Conclusions: These results demonstrate that two CNNs can be used in succession to segment the kidneys with RCC on a coronal CT image with high accuracy. Moreover, as the subjects included patients with RCC, medical renal disease, and cystic renal disease, the outcome further supports the role of CNN in evaluation of complex cases rather than pristine training sets. Future work will focus on segmentation of RCC lesions.