Computer‐aided diagnosis of prostate cancer on magnetic resonance imaging using a convolutional neural network algorithm

• Published in: BJU international Vol. 122; no. 3; pp. 411 - 417
• Main Authors: Ishioka, Junichiro, Matsuoka, Yoh, Uehara, Sho, Yasuda, Yosuke, Kijima, Toshiki, Yoshida, Soichiro, Yokoyama, Minato, Saito, Kazutaka, Kihara, Kazunori, Numao, Noboru, Kimura, Tomo, Kudo, Kosei, Kumazawa, Itsuo, Fujii, Yasuhisa
• Format: Journal Article
• Language: English
• Published: England Wiley Subscription Services, Inc 01.09.2018
• Subjects: Algorithms; Biopsy; computer‐aided diagnosis; Datasets; deep learning; Diagnosis; Magnetic resonance imaging; neural network; Neural networks; NMR; Nuclear magnetic resonance; PCSM; prostate biopsy; Prostate cancer; ProstateCancer; deep learning; prostate biopsy; magnetic resonance imaging; PCSM; ProstateCancer; computer-aided diagnosis; neural network
• ISSN: 1464-4096; 1464-410X; 1464-410X
• DOI: 10.1111/bju.14397

Objective To develop a computer‐aided diagnosis (CAD) algorithm with a deep learning architecture for detecting prostate cancer on magnetic resonance imaging (MRI) to promote global standardisation and diminish variation in the interpretation of prostate MRI. Patients and Methods We retrospectively reviewed data from 335 patients with a prostate‐specific antigen level of <20 ng/mL who underwent MRI and extended systematic prostate biopsy with or without MRI‐targeted biopsy. The data were divided into a training data set (n = 301), which was used to develop the CAD algorithm, and two evaluation data sets (n = 34). A deep convolutional neural network (CNN) was trained using MR images labelled as ‘cancer’ or ‘no cancer’ confirmed by the above‐mentioned biopsy. Using the CAD algorithm that showed the best diagnostic accuracy with the two evaluation data sets, the data set not used for evaluation was analysed, and receiver operating curve analysis was performed. Results Graphics processing unit computing required 5.5 h to learn to analyse 2 million images. The time required for the CAD algorithm to evaluate a new image was 30 ms/image. The two algorithms showed area under the curve values of 0.645 and 0.636, respectively, in the validation data sets. The number of patients mistakenly diagnosed as having cancer was 16/17 patients and seven of 17 patients in the two validation data sets, respectively. Zero and two oversights were found in the two validation data sets, respectively. Conclusion We developed a CAD system using a CNN algorithm for the fully automated detection of prostate cancer using MRI, which has the potential to provide reproducible interpretation and a greater level of standardisation and consistency.