To determine the ability of an AI software tool to detect significant prostate cancer on MRI
To determine the ability of an AI software tool to detect significant prostate cancer on MRI
P Ward, M Sakthithasan, C Thomas, N Trent, DM Koh, PR Burn
To determine the ability of an AI software tool to detect significant prostate cancer on MRI. Prostate cancer is the most common cancer in UK men, with around 52,300 new diagnoses/year, representing 27% of all new cancer cases in males (2016/18). At our institution, referrals have increased by 50% between 2019 and 2022 and <1% of our patients met the UK NHS’s 28 day diagnostic pathway target (MRI and biopsy) for the past year. PiTM (Prostate Intelligence, Lucida Medical) is an AI-based software product for prostate cancer diagnosis on MRI. We evaluated Pi at our institution as a possible aid to improving the workflow of our diagnostic pathway.
Single-centre retrospective study. Participants Consecutive patients were included who had previously been investigated for suspected prostate cancer with MRI, and had either a biopsy result or a radiologist-assigned Likert 1/2 score on MRI (or both). Data Collection All patients were scanned using at 3T (Siemens Vida, 60/200 XT gradients) using a multiparametric protocol. MRIs were analysed using Pi (version 2.2, Lucida Medical). For each patient, the software assigns a Pi score (1.0 to 5.0, ie a 41-point scale) based on increasing probability of significant cancer. Significant prostate cancer was defined as ISUP Grade Group (GG) >/= 2 (Gleason Score 3+4 and above). Data Analysis The median Pi scores of two groups, patients with significant cancer, and those without significant cancer, were compared (Mann-Whitney U test). The median Pi Scores of different GGs were compared (Kruskal-Wallis test). Receiver operating characteristic (ROC) curve analysis was performed for Pi Score diagnostic performance to detect significant cancer. The characteristics of three Pi Score groups were assessed.
Distribution of Pi Scores for No Significant Cancer* and Significant Cancer (≥GG2) Groups *includes Likert 1/2 (no biopsy), benign histology and GG1
Distribution of Pi Scores for Different Patient Groups
ROC Curve for Pi Score (fitted ROC area 0.892)
568 patients were included. Prevalence of significant cancer 45% (257/568). Histology 68% (387/568) of patients had undergone biopsy, of these: benign histology 22% (85/387), GG1 12% (45/387), GG2 26% (99/387), GG3 14% (55/387), GG4/5 27% (103/387). Median Pi Scores (see first two graphs) No significant cancer group, median 3.6 (inter-quartile range 2.9-4.2), significant cancer group, 4.5 (4.2-4.7) (with significant difference between the groups [p < .001]). Likert 1/2 (no biopsy), median 3.2 (2.9-3.7), benign histology 3.9 (3.1-4.2), GG1 4.1 (3.2-4.3), GG2 4.2 (4.1-4.5), GG3 4.6 (4.3-4.8), GG4/5 4.7 (4.5-4.9) (with significant difference between GGs [p 3.5 and <4.5), 41% (100/243) of patients had significant cancer; this group contained 43% (243/568) of all patients. Limitations 32% (181/568) patients did not have biopsy data, for these we have relied on the radiologist’s diagnosis of benignity (Likert Score 1/2); to date, none of these patients have re-presented with evidence of prostate cancer. We excluded patients with a radiologist’s Likert 3 Score who did not have a biopsy. Overall, there were 159 patients with Likert 3 Score, of which 30% (48/159) were excluded.
Proportion of significant cancers in each of three Pi Score groups
Patients with significant cancer have a higher median Pi Score than those without, and a higher median Pi Score is associated with a higher Grade Group on biopsy. Pi Score /= 4.5 has a high PPV for significant prostate cancer. AI tools could be used to help triage MRI findings in patients with suspected prostate cancer and might accelerate diagnostic pathways for prostate cancer.
