Summary
- A 3D anatomical-aware PCa detection network (AtPCa-Net) utilizing anatomical priors was proposed to detect clinically significant PCa (csPCa) from mpMRI images.
- The dataset included 652 patients with and without PCa lesions, and 5-fold cross-validation was used to evaluate the model performance.
- The network architecture incorporates a symmetric-aware network to suppress false positive predictions and a Zonal Loss (ZL) to integrate PCa-related zonal differences into the label and loss design.
- The model was trained on mpMRI images from Siemens 3T scanners using a UNet-like backbone structure and hierarchical label and loss design.
- The AtPCa-Net showed promising results in detecting csPCa lesions utilizing anatomical priors on mpMRI images, offering a potential tool for improving PCa detection in clinical practice.
Doctors and healthcare professionals now have access to a new and advanced tool for detecting clinically significant prostate cancer (csPCa). A recent study introduced a 3D anatomical-aware PCa detection network, known as AtPCa-Net, that utilizes PCa-related anatomical priors to detect csPCa using whole-mount histopathology (WMHP) as confirmation.
The AtPCa-Net consists of two main parts. Firstly, a 3D symmetric-aware network considers symmetric-related information to reduce false positive predictions. Secondly, the Zonal Loss (ZL) part integrates PCa-related zonal differences into the label and loss design. The overall architecture of the AtPCa-Net is based on the nnU-Net structure, known for its performance in detection and segmentation of medical imaging tasks.
The dataset used in the study included mpMRI images from 652 patients with confirmed PCa lesions and negative findings. The images were reviewed by experienced genitourinary radiologists and pathologists to ensure accurate annotations and ground truth for the study.
One key aspect of the study was the implementation of a hierarchical label and loss design, the ZL, which accounted for the zonal appearance differences of PCa lesions in different zones of the prostate. This design aimed to guide the model in learning the distinct appearances of PCa lesions in different zones with anatomical constraints.
The network architecture of the AtPCa-Net included shared-weight encoders to maintain symmetric features, resembling the human visual system, and enhance the model’s decision-making process. The output of the network provides a probability map of where suspicious csPCa lesions are located.
In conclusion, the AtPCa-Net offers a promising approach to improving the detection of clinically significant prostate cancer using advanced 3D imaging technology and anatomical-aware priors. This innovative tool has the potential to enhance diagnostic accuracy and improve patient outcomes in the field of prostate cancer detection.
Radiology,Urology