한빛사
>
한빛사논문
한빛사논문
- 조회67
Radiology, Published Online:Mar 11 2025 | https://doi.org/10.1148/radiol.232788
MRI-based Deep Learning Algorithm for Assisting Clinically Significant Prostate Cancer Detection: A Bicenter Prospective Study
Young Joon Lee*,1, Hyong Woo Moon*,2, Moon Hyung Choi1, Seung Eun Jung1, Yong Hyun Park2, Ji Youl Lee2, Dong Hwan Kim3, Sung Eun Rha3, Sang Hoon Kim4, Kyu Won Lee4, Yeong-Jin Choi5, Young Sub Lee6, Woojoo Lee7,8, Seungjae Lee8, Robert Grimm9, Heinrich von Busch9, Dongyeob Han10, Bin Lou11, Ali Kamen11
1Department of Radiology, Eunpyeong St Mary’s Hospital, College of Medicine, The Catholic University of Korea, 1021 Tongil-ro, Eunpyeong-gu, Seoul 03312, Republic of Korea
2Department of Urology, Seoul St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
3Department of Radiology, Seoul St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
4Department of Urology, Eunpyeong St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
5Department of Hospital Pathology, Seoul St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
6Department of Hospital Pathology, Eunpyeong St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
7Department of Public Health Sciences, Graduate School of Public Health, Seoul National University, Seoul, Republic of Korea
8Institute of Health and Environment, Seoul National University, Seoul, Republic of Korea
9Diagnostic Imaging, Siemens Healthineers, Forchheim, Germany
10Diagnostic Imaging, Siemens Healthineers, Seoul, Republic of Korea
11Digital Technology and Innovation, Siemens Healthineers, Princeton, NJ
* Y.J.L. and H.W.M. contributed equally to this work.
Address correspondence to: M.H.C.
1Department of Radiology, Eunpyeong St Mary’s Hospital, College of Medicine, The Catholic University of Korea, 1021 Tongil-ro, Eunpyeong-gu, Seoul 03312, Republic of Korea
2Department of Urology, Seoul St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
3Department of Radiology, Seoul St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
4Department of Urology, Eunpyeong St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
5Department of Hospital Pathology, Seoul St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
6Department of Hospital Pathology, Eunpyeong St Mary’s Hospital, College of Medicine, The Catholic University of Korea, Seoul, Republic of Korea
7Department of Public Health Sciences, Graduate School of Public Health, Seoul National University, Seoul, Republic of Korea
8Institute of Health and Environment, Seoul National University, Seoul, Republic of Korea
9Diagnostic Imaging, Siemens Healthineers, Forchheim, Germany
10Diagnostic Imaging, Siemens Healthineers, Seoul, Republic of Korea
11Digital Technology and Innovation, Siemens Healthineers, Princeton, NJ
* Y.J.L. and H.W.M. contributed equally to this work.
Address correspondence to: M.H.C.
Abstract
Background
Although artificial intelligence is actively being developed for prostate MRI, few studies have prospectively validated these tools.
Purpose
To compare the diagnostic performance of a commercial deep learning algorithm (DLA) and radiologists’ clinical reports for cancer detection in participants from two hospitals using histopathologic findings from biopsy specimens as the reference standard.
Materials and Methods
This prospective bicenter study enrolled participants suspected of having prostate cancer (PCa) who were scheduled for biopsy based on clinical information, including prostate MRI, from December 2022 to July 2023. Targeted prostate biopsies were performed for lesions with Prostate Imaging Reporting and Data System (PI-RADS) scores of 3 or higher, identified by either radiologists or the DLA. PI-RADS classifications by radiologists (using all imaging sequences), the DLA (using biparametric MRI), and the scenario in which radiologist-based PI-RADS 3 scores were modulated with DLA-based PI-RADS scores were compared using the area under the receiver operating characteristic curve (AUC) with DeLong and McNemar tests.
Results
A total of 259 lesions, including 117 clinically significant PCas (csPCas) (Gleason grade group ≥2), were evaluated in 205 men (median age, 68 years; age range, 47–90 years). At per-lesion analysis, the DLA had a lower sensitivity (94 of 117; 80%) and higher positive predictive value (PPV) (94 of 163; 58%) for detecting csPCa than did the radiologists (109 of 117 [93%] and 109 of 227 [48%]; P = .02 and P = .008, respectively). At per-participant analysis, incorporation of the DLA increased specificity from 23 of 108 (21%) to 48 of 108 (44%) (P = .001), with similar sensitivity (96 of 97 [99%] vs 93 of 97 [96%]; P = .74). There was no evidence of a difference in the AUC between radiologist-based and DLA-based PI-RADS score (0.77 [95% CI: 0.70, 0.82] vs 0.79 [95% CI: 0.73, 0.85]; P = .73).
Conclusion
The DLA demonstrated lower sensitivity but a greater PPV than radiologists for detecting csPCa in a biopsy setting. Using DLA results when radiologists’ interpretations are indeterminate could improve specificity while maintaining sensitivity.
Although artificial intelligence is actively being developed for prostate MRI, few studies have prospectively validated these tools.
Purpose
To compare the diagnostic performance of a commercial deep learning algorithm (DLA) and radiologists’ clinical reports for cancer detection in participants from two hospitals using histopathologic findings from biopsy specimens as the reference standard.
Materials and Methods
This prospective bicenter study enrolled participants suspected of having prostate cancer (PCa) who were scheduled for biopsy based on clinical information, including prostate MRI, from December 2022 to July 2023. Targeted prostate biopsies were performed for lesions with Prostate Imaging Reporting and Data System (PI-RADS) scores of 3 or higher, identified by either radiologists or the DLA. PI-RADS classifications by radiologists (using all imaging sequences), the DLA (using biparametric MRI), and the scenario in which radiologist-based PI-RADS 3 scores were modulated with DLA-based PI-RADS scores were compared using the area under the receiver operating characteristic curve (AUC) with DeLong and McNemar tests.
Results
A total of 259 lesions, including 117 clinically significant PCas (csPCas) (Gleason grade group ≥2), were evaluated in 205 men (median age, 68 years; age range, 47–90 years). At per-lesion analysis, the DLA had a lower sensitivity (94 of 117; 80%) and higher positive predictive value (PPV) (94 of 163; 58%) for detecting csPCa than did the radiologists (109 of 117 [93%] and 109 of 227 [48%]; P = .02 and P = .008, respectively). At per-participant analysis, incorporation of the DLA increased specificity from 23 of 108 (21%) to 48 of 108 (44%) (P = .001), with similar sensitivity (96 of 97 [99%] vs 93 of 97 [96%]; P = .74). There was no evidence of a difference in the AUC between radiologist-based and DLA-based PI-RADS score (0.77 [95% CI: 0.70, 0.82] vs 0.79 [95% CI: 0.73, 0.85]; P = .73).
Conclusion
The DLA demonstrated lower sensitivity but a greater PPV than radiologists for detecting csPCa in a biopsy setting. Using DLA results when radiologists’ interpretations are indeterminate could improve specificity while maintaining sensitivity.
논문정보
- Research article
- 2025년 03월 (BRIC 등록일 2025-03-27)
- 국내(교신)+국외 연구진
댓글 작성 로그인
팝업 닫기관련 링크
연구자 키워드
연구자 ID
관련분야 연구자보기
소속기관 논문보기
관련분야 논문보기
해당논문 저자보기
