A pilot study of AI-assisted reading of prostate MRI in Organized Prostate Cancer Testing
(2024) In Acta Oncologica 63. p.816-821- Abstract
Objectives: To evaluate the feasibility of AI-assisted reading of prostate magnetic resonance imaging (MRI) in Organized Prostate cancer Testing (OPT). Methods: Retrospective cohort study including 57 men with elevated prostate-specific antigen (PSA) levels ≥3 µg/L that performed bi-parametric MRI in OPT. The results of a CE-marked deep learning (DL) algorithm for prostate MRI lesion detection were compared with assessments performed by on-site radiologists and reference radiologists. Per patient PI-RADS (Prostate Imaging-Reporting and Data System)/Likert scores were cross-tabulated and compared with biopsy outcomes, if performed. Positive MRI was defined as PI-RADS/Likert ≥4. Reader variability was assessed with weighted kappa scores.... (More)
Objectives: To evaluate the feasibility of AI-assisted reading of prostate magnetic resonance imaging (MRI) in Organized Prostate cancer Testing (OPT). Methods: Retrospective cohort study including 57 men with elevated prostate-specific antigen (PSA) levels ≥3 µg/L that performed bi-parametric MRI in OPT. The results of a CE-marked deep learning (DL) algorithm for prostate MRI lesion detection were compared with assessments performed by on-site radiologists and reference radiologists. Per patient PI-RADS (Prostate Imaging-Reporting and Data System)/Likert scores were cross-tabulated and compared with biopsy outcomes, if performed. Positive MRI was defined as PI-RADS/Likert ≥4. Reader variability was assessed with weighted kappa scores. Results: The number of positive MRIs was 13 (23%), 8 (14%), and 29 (51%) for the local radiologists, expert consensus, and DL, respectively. Kappa scores were moderate for local radiologists versus expert consensus 0.55 (95% confidence interval [CI]: 0.37–0.74), slight for local radiologists versus DL 0.12 (95% CI: −0.07 to 0.32), and slight for expert consensus versus DL 0.17 (95% CI: −0.01 to 0.35). Out of 10 cases with biopsy proven prostate cancer with Gleason ≥3+4 the DL scored 7 as Likert ≥4. Interpretation: The Dl-algorithm showed low agreement with both local and expert radiologists. Training and validation of DL-algorithms in specific screening cohorts is essential before introduction in organized testing.
(Less)
- author
- Thimansson, Erik ; Zackrisson, Sophia LU ; Jäderling, Fredrik ; Alterbeck, Max LU ; Jiborn, Thomas LU ; Bjartell, Anders LU and Wallström, Jonas
- organization
-
- Radiology Diagnostics, Malmö (research group)
- LUCC: Lund University Cancer Centre
- EpiHealth: Epidemiology for Health
- LTH Profile Area: Photon Science and Technology
- LU Profile Area: Light and Materials
- Urological cancer, Malmö (research group)
- eSSENCE: The e-Science Collaboration
- Division of Translational Cancer Research
- publishing date
- 2024
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- artificial intelligence, Magnetic resonance imaging, overdiagnosis, prostatespecific antigen, prostatic neoplasms
- in
- Acta Oncologica
- volume
- 63
- pages
- 6 pages
- publisher
- Taylor & Francis
- external identifiers
-
- pmid:39473176
- scopus:85208167411
- ISSN
- 0284-186X
- DOI
- 10.2340/1651-226X.2024.40475
- language
- English
- LU publication?
- yes
- id
- fb93a52b-01c5-4fc1-87db-e60907ab85a3
- date added to LUP
- 2024-12-10 13:51:35
- date last changed
- 2025-07-09 06:59:06
@article{fb93a52b-01c5-4fc1-87db-e60907ab85a3, abstract = {{<p>Objectives: To evaluate the feasibility of AI-assisted reading of prostate magnetic resonance imaging (MRI) in Organized Prostate cancer Testing (OPT). Methods: Retrospective cohort study including 57 men with elevated prostate-specific antigen (PSA) levels ≥3 µg/L that performed bi-parametric MRI in OPT. The results of a CE-marked deep learning (DL) algorithm for prostate MRI lesion detection were compared with assessments performed by on-site radiologists and reference radiologists. Per patient PI-RADS (Prostate Imaging-Reporting and Data System)/Likert scores were cross-tabulated and compared with biopsy outcomes, if performed. Positive MRI was defined as PI-RADS/Likert ≥4. Reader variability was assessed with weighted kappa scores. Results: The number of positive MRIs was 13 (23%), 8 (14%), and 29 (51%) for the local radiologists, expert consensus, and DL, respectively. Kappa scores were moderate for local radiologists versus expert consensus 0.55 (95% confidence interval [CI]: 0.37–0.74), slight for local radiologists versus DL 0.12 (95% CI: −0.07 to 0.32), and slight for expert consensus versus DL 0.17 (95% CI: −0.01 to 0.35). Out of 10 cases with biopsy proven prostate cancer with Gleason ≥3+4 the DL scored 7 as Likert ≥4. Interpretation: The Dl-algorithm showed low agreement with both local and expert radiologists. Training and validation of DL-algorithms in specific screening cohorts is essential before introduction in organized testing.</p>}}, author = {{Thimansson, Erik and Zackrisson, Sophia and Jäderling, Fredrik and Alterbeck, Max and Jiborn, Thomas and Bjartell, Anders and Wallström, Jonas}}, issn = {{0284-186X}}, keywords = {{artificial intelligence; Magnetic resonance imaging; overdiagnosis; prostatespecific antigen; prostatic neoplasms}}, language = {{eng}}, pages = {{816--821}}, publisher = {{Taylor & Francis}}, series = {{Acta Oncologica}}, title = {{A pilot study of AI-assisted reading of prostate MRI in Organized Prostate Cancer Testing}}, url = {{http://dx.doi.org/10.2340/1651-226X.2024.40475}}, doi = {{10.2340/1651-226X.2024.40475}}, volume = {{63}}, year = {{2024}}, }