Presentation Authors: Junichiro Ishioka*, Tokyo, Japan, Masaki Kobayashi, Tsuchiura, Japan, Motohiro Fujiwara, Naoko Kawamura, Tetsuo Okuno, Toride, Japan, Yuichi Fukuda, Tomoaki Kohno, Keizo Kawano, Shinji Morimoto, Tsuchiura, Japan, Sho Uehara, Yosuke Yasuda, Toshiki Kijima, Soichiro Yoshida, Minato Yokoyama, Yoh Matsuoka, Kazutaka Saito, Ryota Saiki, Itsuo Kumazawa, Yasuhisa Fujii, Tokyo, Japan
Introduction: Computed tomography (CT) improves the detection of urinary tract stones (UTS). However, availability, quality, and cost differ depending on the country and some patients cannot undergo the examination because of radiation exposure concerns. The KUB is the lowest cost and the most widely available examination but its low reliability is a drawback. To improve the diagnostic performance of KUB, we used computer-aided diagnosis with a convolutional neural network (CNN) algorithm.
Methods: We used 1017 KUB images of patients who were diagnosed with UTS at three hospitals. The existence of radio-opaque UTS was confirmed by CT or follow-up KUB indicating stone discharge. Each image was examined by urologists. Overall, 827 images from two hospitals were used as training data and 190 images from the remaining hospital were used as test data to assess performance. We cropped parallelogram images containing stones and performed an affine transformation to create a square of 166 Ã— 166 pixels. Deep learning for parameter optimization was performed using a CNN (ResNet). The ResNet has architecture that uses skip connections and short-cuts to jump over layers.
Results: The parameter optimization time was 9 hours and the required time for detecting the UTS using a trained algorithm was 110 ms (GPU: NVIDIA GeForceGTX1080Â®). In the test dataset, the positive predictive value, sensitivity, and F-measure were 0.49, 0.72, and 0.58, respectively.
Conclusions: We developed a computer-aided diagnostic system using convolutional neural network to detect radio-opaque urinary tract stones in KUB, which could provide a reproducible interpretation and a greater level of standardization and consistency.