Clinical usefulness of deep learning-based automated segmentation in intracranial hemorrhage

Article type: Research Article

Authors: Kim, Chang Ho^{a; 1} | Hahm, Myong Hun^{b; 1} | Lee, Dong Eun^a | Choe, Jae Young^a | Ahn, Jae Yun^a | Park, Sin-Youl^c | Lee, Suk Hee^d | Kwak, Youngseok^e | Yoon, Sang-Youl^e | Kim, Ki-Hong^f | Kim, Myungsoo^e | Chang, Sung Hyun^e | Son, Jeongwoo^c | Cho, Junghwan^g | Park, Ki-Su^{e; *} | Kim, Jong Kun^{a; *}

Affiliations: [a] Department of Emergency Medicine, School of Medicine, Kyungpook National University, Daegu, Korea | [b] Department of Radiology, School of Medicine, Kyungpook National University, Daegu, Korea | [c] Department of Emergency Medicine College of Medicine, Yeungnam University, Daegu, Korea | [d] Department of Emergency Medicine Daegu Catholic University Medical Center, Daegu, Korea | [e] Department of Neurosurgery, School of Medicine, Kyungpook National University, Daegu, Korea | [f] Department of Neurosurgery, School of Medicine of Daegu Catholic University, Daegu, Korea | [g] CAIDE Systems Inc., USA

Correspondence: [*] Corresponding authors: Ki-Su Park, Department of Neurosurgery, School of Medicine, Kyungpook National University, 130 Dongdeok-ro, Jung-gu, Daegu 41944, Korea. E-mail: [email protected]. Jong Kun Kim, Department of Emergency Medicine, School of Medicine, Kyungpook National University, 130 Dongdeok-ro, Jung-gu, Daegu 41944, Korea. E-mail: [email protected].

Note: [1] Chang Ho Kim and Myong Hun Hahm contributed equally to this work.

Abstract: BACKGROUND:Doctors with various specializations and experience order brain computed tomography (CT) to rule out intracranial hemorrhage (ICH). Advanced artificial intelligence (AI) can discriminate subtypes of ICH with high accuracy. OBJECTIVE: The purpose of this study was to investigate the clinical usefulness of AI in ICH detection for doctors across a variety of specialties and backgrounds. METHODS: A total of 5702 patients’ brain CTs were used to develop a cascaded deep-learning-based automated segmentation algorithm (CDLA). A total of 38 doctors were recruited for testing and categorized into nine groups. Diagnostic time and accuracy were evaluated for doctors with and without assistance from the CDLA. RESULTS: The CDLA in the validation set for differential diagnoses among a negative finding and five subtypes of ICH revealed an AUC of 0.966 (95% CI, 0.955–0.977). Specific doctor groups, such as interns, internal medicine, pediatrics, and emergency junior residents, showed significant improvement with assistance from the CDLA (p= 0.029). However, the CDLA did not show a reduction in the mean diagnostic time. CONCLUSIONS: Even though the CDLA may not reduce diagnostic time for ICH detection, unlike our expectation, it can play a role in improving diagnostic accuracy in specific doctor groups.

Keywords: Intracranial hemorrhages, diagnosis, artificial intelligence, deep learning, ROC curve

DOI: 10.3233/THC-202533

Journal: Technology and Health Care, vol. 29, no. 5, pp. 881-895, 2021