CT prostate segmentation based on synthetic MRI-aided deep attention fully convolution network.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Prostate segmentation on CT scans is difficult due to poor soft tissue contrast.
• MRI aids delineation but is limited by registration errors between MRI and CT.
• Existing methods struggle with accurate prostate segmentation on CT alone.

Purpose of the Study:

• To develop a CT-only prostate segmentation strategy using synthetic MRI (sMRI).
• To improve the accuracy of prostate delineation for treatment planning without MRI acquisition.
• To overcome the limitations of CT contrast and MRI-CT registration errors.

Main Methods:

• A deep attention-based network was developed for prostate segmentation on CT.
• Synthetic MRI (sMRI) was generated from CT images using a cycle generative adversarial network.
• An attention fully convolutional network, trained on sMRI and deformed MRI contours, was used for segmentation.

Main Results:

• The method achieved high segmentation accuracy, with Dice similarity coefficients of 0.92 ± 0.09 (leave-one-out) and 0.91 ± 0.07 (hold-out test).
• Quantitative metrics including Hausdorff distance and mean surface distance demonstrated the technique's precision.
• Validation was performed on a total of 99 patients, confirming robust performance.

Conclusions:

• A novel CT-only prostate segmentation strategy using CT-based sMRI was successfully developed.
• The technique provides accurate prostate segmentation without the need for MRI acquisition.
• This approach facilitates routine clinical workflows by simplifying treatment planning.