[2P1-GS-10] AI application: medicine

[2P1-GS-10-01] Kidney Cancer Detection Based on Multi-Phase Convolutional Neural Networks and Synthetic Contrast-Enhanced CT Images

Junsei Suzuki¹, Rei Kuramoto¹, 〇Yoshitaka Kameya¹, Keiichi Yamada¹, Kazuhiro Hotta¹, Tomoichi Takahashi², Naoto Sassa³, Yoshihisa Matsukawa⁴, Shingo Iwano⁴, Tokunori Yamamoto⁴ (1. Meijo University, 2. Meis Technology Inc., 3. Aichi Medical University, 4. Nagoya University)

[[Online]]

Keywords:Kidney Cancer Detection, Contrast-Enhanced CT Images, Image-to-Image Translation, Convolutional Neural Networks

This study focuses on convolutional neural networks (CNNs) that detect kidney cancer from computed tomography (CT) images. Such CNNs would enable us to screen patients prior to diagnosis by medical experts. CT images obtained under the administration of contrast agents are known as contrast-enhanced CT (CECT) images. While contrast agents enhance the contrast among tissues, diagnosis with unenhanced CT (UCT) images is preferred if patients are allergic to contrast agents or contrast agents worsen the renal functions of patients. To detect kidney cancer without contrast agents, in this study, we propose to generate synthetic CECT images based on the existing aligned pairs of UCT and CECT images. In generating such synthetic CECT images, we use pix2pix, a well-known generative adversarial network for image-to-image transformation. Moreover, for robust detection, we introduce multi-phase CNNs that receive UCT and CECT images simultaneously. The experimental results show that multi-phase CNNs with real UCT and synthetic CECT images perform better than conventional CNNs with real UCT images.