Accurate localisation of prostate lesions is essential for the effective diagnosis and management of prostate cancer, which remains one of the most common cancers in men across Europe and North America. Multiparametric MRI (mpMRI) plays a crucial role in the detection of clinically significant prostate cancer, offering benefits such as reducing overdiagnosis and limiting unnecessary treatment. However, consistently accurate reporting remains a significant challenge, particularly with the Prostate Imaging Reporting and Data System (PI-RADS), which requires the use of a 41-sector map that can be time-consuming and complex. To address this, a deep learning-based segmentation algorithm featuring a simplified 24-sector grid map has been developed. This innovation is designed to enhance lesion localisation by improving precision and usability while supporting the demands of high-volume clinical settings.
A Streamlined Approach to Sector Mapping
The 24-sector grid map was designed through the development of an automatic segmentation grid (ASG) that uses a deep learning algorithm, based on the nnU-Net model. This algorithm was trained on a large dataset that included axial T2-weighted, ADC, and DWI images, supported by manually annotated segmentation data. With strong performance demonstrated through high Dice coefficients for both whole gland and transition zone segmentation, the model provides a robust foundation for accurate mapping.
Must Read: Low-Field MRI in Prostate Cancer Screening: A Viable Alternative to 3T?
The segmentation approach divides the prostate gland into base, mid-gland, and apex sections, and further subdivides these into anterior and posterior regions, as well as left and right sides, resulting in 24 total sectors. This design aligns with commonly used clinical descriptors, making it more intuitive for radiologists. The algorithm is implemented using Python and standard libraries, and once integrated into imaging systems, it can superimpose the grid onto T2-weighted images in approximately one second, allowing for efficient and repeatable sector assignment during routine practice.
Performance Compared to Radiologists
The effectiveness of the ASG algorithm was assessed in a retrospective study that included 50 patients who had previously undergone mpMRI examinations. Fiducials representing random lesion locations were placed in both the peripheral and transition zones of the prostate and reviewed by four independent radiologists. Each radiologist first identified lesion locations without the support of the sector map, and later repeated the task with the map superimposed. In addition, the ASG algorithm independently assigned lesion locations. Without the assistance of the grid, radiologists correctly identified the sector in 55% of cases. When supported by the grid map, the accuracy increased to 71%. By comparison, the ASG algorithm achieved an 80% accuracy rate in identifying the correct sector.
The use of the grid map also led to a notable reduction in assignments to adjacent or incorrect sectors by radiologists, while the algorithm consistently showed fewer misplacements. These findings were particularly marked in the transition zone, where radiologist accuracy without assistance was significantly lower. Statistical analysis demonstrated a significant improvement in performance when using the grid map and confirmed that the algorithm's accuracy was not statistically different from that of radiologists assisted by the map.
Clinical Implications and Limitations
The ASG algorithm provides a practical solution to improve the reproducibility and accuracy of prostate lesion localisation on mpMRI. In clinical practice, especially in busy or resource-limited environments, the ability to automate and standardise this aspect of imaging could enhance diagnostic confidence and support multidisciplinary communication. The tool’s utility extends to follow-up imaging, where consistent lesion localisation is essential, particularly in cases under active surveillance. The algorithm’s speed and ease of integration into structured reporting systems also make it a suitable candidate for broader adoption.
However, the study did acknowledge some limitations. The use of small, random fiducials does not fully replicate real clinical lesions, which may span multiple sectors. Additionally, the tool does not currently support mapping of lesions extending beyond the prostate, such as into the seminal vesicles or adjacent organs. Furthermore, as the study was conducted at a single institution, further research is needed to confirm these results across varied clinical settings and patient populations.
The study confirmed the potential of artificial intelligence, and in particular deep learning algorithms, in improving mpMRI-based prostate lesion localisation through an automated 24-sector grid map. The ASG algorithm demonstrated a higher rate of correct lesion localisation than unassisted radiologists and also contributed to improved accuracy when used as a support tool. The proposed 24-sector map strikes a balance between complexity and usability, making it a practical tool for enhancing structured reporting in prostate imaging. With the potential for rapid deployment in high-volume clinical workflows, this tool could improve diagnostic reproducibility and streamline communication between radiologists and referring clinicians. Future multicentre studies will be essential in confirming its effectiveness and supporting widespread clinical integration.
Source: European Journal of Radiology
Image Credit: iStock
