The use of artificial intelligence (AI) to diagnose cancer and distinguish prostate cancers from benign tumors has been extensively investigated. However, existing deep learning methods for diagnosing prostate cancer are based on magnetic resonance imaging (MRI) data alone, and diagnosis relies on ground truth data determined by radiologists. Therefore, current deep learning algorithms provide limited accuracy in predicting the 3D mapping of prostate tumors.
Researchers from the Kyoto Prefectural University of Medicine developed an AI model to predict the location and volume of clinically significant prostate cancer by integrating multiparametric magnetic resonance-ultrasound (MR-US) image data and fusion biopsy trajectory-proven pathology data.1
In a proof-of-concept study, the researchers showed that this pilot AI algorithm was more precise than a radiologist’s reading in predicting the location and volume of clinically significant prostate cancers. By predicting the 3D mapping of clinically significant prostate cancer, this novel AI model may help in the planning of focal therapy to achieve more precise and complete ablation of tumors in patients with prostate cancer.1
The study was published in the journal The Prostate on February 22, 2022.
Developing an AI Algorithm Using Multiparametric Magnetic Resonance Imaging Data
To predict the 3D mapping of prostate tumors, Dr. Kaneko and colleagues integrated multiparametric MR-US image data and fusion biopsy trajectory-proven pathology data to develop an AI algorithm to predict the volume and location of clinically significant cancer.
The team collected pathological data from 20 consecutive patients who underwent MRI-US fusion prostate biopsy, followed by robot-assisted radical prostatectomy. Using multiparametric MR-US image data and MRI-US prostate biopsy trajectory-proven pathology data (compensated on 16‐channel images), they trained a convolutional neural network (CNN) to predict the 3D mapping of prostate tumors. For machine learning, they used 15 MRI parameters and 22,968 histologically labeled voxel data from 171 biopsy trajectory images.
After training the algorithm using 18,372 random patch images (80% of the labeled patch images), the researchers applied the algorithm to 4,596 patch images (validation set) to generate cancer prediction maps. AI-predicted 3D maps were overlaid on multiparametric MR images and compared with tumors from patients who underwent robot-assisted radical prostatectomy.
AI Predicts the Location of Clinically Significant Prostate Cancer
To determine the ability of the pilot AI algorithm to predict the location of prostate tumors, Dr. Kaneko and coworkers compared the AI-assisted 3D mapping of robot-assisted radical prostatectomy specimens with the 3D mapping of experienced radiologists. Radiologists performed 3D mapping of prostate tumors using MRI data and the Prostate Imaging Reporting and Data System (PI-RADS) version 2.
The AI-assisted 3D mapping provided a high concordance (83%) of the centers of clinically relevant prostate tumors with those in radical prostatectomy specimens. In comparison, the radiologist’s reading provided a 54% concordance of the center of prostate tumors with that in prostatectomy specimens; this difference between the two methods was statistically significant (p = 0.036). These findings suggest that AI can more precisely predict the location and distribution of clinically significant prostate cancer than experienced radiologists’ readings.
Compared with the radiologist’s reading, the AI algorithm provided higher detection rates for any cancer (46% vs. 32%), clinically significant prostate cancer (76% vs. 63%), index cancer (100% vs. 95%), clinically insignificant prostate cancer (18% vs. 2.5%), and low‐volume (defined as ≤0.5 cm3) clinically significant prostate cancer (50% vs. 28%). However, the differences in these predictions between the two methods were not statistically significant.
AI Predicts the Volume of Clinically Significant Prostate Cancer
To determine the accuracy of the pilot AI algorithm in predicting the volume of clinically significant prostate tumors, the team assessed the AI-predicted volumes of prostate tumors and those measured by experienced radiologists based on MRI data. Compared with prostate tumor volumes based on radiologists’ readings, those predicted with AI were more accurate. Assessment of the relationship between AI and histological volumes of clinically significant prostate cancer revealed a high correlation (Pearson’s correlation r = 0.90, p < 0.001).
By increasing the accuracy of prediction of the volume of clinically significant prostate cancer, AI may help decrease the treatment margin and reduce the burden of treatment‐related comorbidities.
Challenges and Future Perspectives
Although the new AI algorithm could more precisely predict the location and volume of clinically significant prostate cancer than radiologists’ readings, the algorithm was inferior to radiologists’ readings in false positive diagnosis (false positive rate: 7.4% for the radiologist’s reading vs. 29% for AI, p = 0.041), raising concerns about potential overtreatment.
Moreover, the algorithm failed to identify lesions with PI-RADS score ≥2 as suspicious prostate tumors. AI missed nine clinically significant prostate tumors in six patients. All nine of the undetected tumors were Gleason grade 2–4 tumors. Small (< 0.5 cm3) clinically significant prostate tumors and those with low PI‐RADS version 2 scores were more likely to be missed by the AI algorithm. Improving ground truth data and the integration of digitally recorded and actual trajectory data may help reduce the false positive rate of the algorithm. The sensitivity of the AI algorithm should also be further improved to ensure the detection of small prostate tumors.
As none of the biopsies were tumors with centers in areas other than the peripheral zone or transition zone, the AI algorithm was limited to predicting cancers with centers in peripheral zone or transition zone segmentation. Incorporating malignant trajectory data from other areas is required to enable the algorithm to predict cancer with a center in other areas.
Additionally, the study cohort was small and received treatment at a single institution; therefore, future multicenter studies are required to confirm the ability of this new AI algorithm to predict the location and volume of clinically significant prostate cancers. Another limitation of the AI algorithm is that it is not fully automated, as transition zone/peripheral zone segmentation and the evaluation of cancer location in biopsy cores must be performed manually by a pathologist.
References
- Kaneko M, Fukuda N, Nagano H, et al. Artificial intelligence trained with integration of multiparametric MR-US imaging data and fusion biopsy trajectory-proven pathology data for 3D prediction of prostate cancer: A proof-of-concept study. Prostate. 2022;n/a(n/a). doi:https://doi.org/10.1002/pros.24321
