AtPCa-Net: anatomical-aware prostate cancer detection network on multi-parametric MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Multi-parametric MRI (mpMRI) is crucial for prostate cancer (PCa) diagnosis.
• Current deep learning models may overlook domain-specific anatomical information in mpMRI, potentially limiting PCa detection performance.
• Prostate anatomy, including symmetry and zonal location, is vital for differentiating PCa from benign conditions and assessing aggressiveness.

Purpose of the Study:

• To investigate the impact of domain-specific anatomical properties on PCa diagnosis using mpMRI.
• To develop and evaluate an anatomical-aware deep learning framework to improve PCa detection.
• To enhance the utilization of anatomical information within deep learning models for PCa diagnosis.

Main Methods:

• Proposed an anatomical-aware PCa detection Network (AtPCa-Net) integrating domain-specific anatomical features.
• Trained and tested AtPCa-Net on mpMRI datasets for PCa detection.
• Compared the performance of AtPCa-Net against conventional deep learning approaches.

Main Results:

• AtPCa-Net demonstrated improved utilization of anatomical information compared to standard models.
• The proposed anatomical-aware designs led to enhanced overall model performance.
• Significant improvements were observed in both PCa lesion detection and patient-level classification accuracy.

Conclusions:

• Incorporating domain-specific anatomical knowledge into deep learning models significantly boosts PCa detection on mpMRI.
• AtPCa-Net offers a promising approach for more accurate and reliable PCa diagnosis.
• Future research should focus on further refining anatomical feature integration in AI for medical imaging.