AI prostate cancer detection is a rapidly evolving field, yet recent studies highlight key limitations when compared to human expertise. A new study evaluated a deep-learning segmentation model (nnU-Net) in a prostate MRI screening population, comparing its performance against radiologists using the established PI-RADS v2 metrics. Consequently, the research provides crucial data regarding the current utility of artificial intelligence in a real-world clinical setting, especially for identifying clinically significant prostate cancer (csPC).
Study Design and Methodology
Researchers curated a prostate MRI dataset from a cancer screening population known as the G2-trial. The dataset was divided into a training set (1086 examinations) and a test set (268 examinations) from 1254 men. Histopathology served as the definitive reference standard. Furthermore, to enhance the reliability of negative findings, the protocol mandated that 288 men underwent systematic biopsies regardless of their MRI results. Moreover, all men were monitored with a minimum of three years of follow-up. Since the goal was to detect csPC, it was defined as International Society of Urological Pathology (ISUP) grade 2 or higher.
AI Prostate Cancer Performance Metrics
The AI system achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.83 (95% CI 0.73–0.92) in the test cohort. However, a significant finding emerged when comparing the AI’s diagnostic ability to that of the human readers. While the AI demonstrated a certain level of sensitivity, it presented significantly lower specificity compared to radiologists at matched sensitivity levels. This implies the neural network generated a higher number of false positive results, a critical issue in screening programs.
Context and Clinical Relevance for Indian Practice
Multiparametric MRI (mp-MRI) is a cornerstone of modern prostate cancer diagnosis, especially in patients with elevated PSA and a negative DRE, as per Indian guidelines. The Indian Council of Medical Research (ICMR) recommends mp-MRI prior to systematic biopsy, particularly if the PI-RADS score is 3 or greater. Moreover, AI is transforming diagnosis by automating tasks like segmentation and lesion detection, potentially improving efficiency for radiologists globally. In a different large-scale study, AI assistance actually increased diagnostic accuracy, particularly for non-expert readers. Therefore, the goal for future AI development must be to maintain high sensitivity while improving specificity to minimise unnecessary follow-up procedures like biopsies. Ultimately, this lower specificity in a screening cohort suggests AI is not yet ready for autonomous use.
Frequently Asked Questions
Q1: How did the AI system compare to radiologists in the screening study?
The AI system achieved an AUROC of 0.83 but demonstrated significantly lower specificity compared to radiologists at matched sensitivity levels. This means the AI produced more false-positive results than the human readers using the PI-RADS v2 system.
Q2: What is the clinical relevance of the AI’s low specificity?
A lower specificity increases the false-positive rate. Consequently, this leads to unnecessary anxiety, costly further investigations, and potentially invasive procedures, such as a biopsy, for men who do not have clinically significant prostate cancer.
Q3: What is clinically significant prostate cancer (csPC) in this context?
The study defined csPC as International Society of Urological Pathology (ISUP) grade 2 or higher, which typically corresponds to a Gleason score of 3+4 or greater.
References
- Langkilde F et al. Evaluation of AI for prostate cancer detection in biparametric-MRI screening population data. Eur Radiol. 2025 Dec 08. doi: 10.1007/s00330-025-12198-5. PMID: 41359160.
- AI model developed to detect prostate cancer found to perform like ‘experienced radiologist’. indiatimes.com.
- Artificial Intelligence Enhances Prostate Cancer Detection in MRI Study. Medscape.
- Consensus Document for Management of Prostate Cancer. icmr.gov.in.