Fully automatic reconstruction of prostate high-dose-rate brachytherapy interstitial needles using two-phase deep...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate localization of brachytherapy (BT) applicator/needle trajectories is crucial for successful treatment.
• Current methods for detecting needle paths in prostate BT can be challenging and time-consuming.

Purpose of the Study:

• To develop and validate a two-phase deep learning approach for automated localization of high-dose-rate (HDR) prostate BT catheters.
• To enhance the accuracy and efficiency of needle trajectory detection in prostate brachytherapy.

Main Methods:

• A two-phase deep learning model was employed, utilizing a pix2pix Generative Adversarial Network (GAN) for needle segmentation and Generic Object Tracking Using Regression Networks (GOTURN) for trajectory prediction.
• The models were trained and tested on a clinical dataset of 25 patients undergoing prostate HDR-BT, with 5 patients reserved for testing.

Main Results:

• The pix2pix GAN achieved 98.72% segmentation accuracy for needles.
• High performance metrics were reported: Dice Similarity Coefficient (DSC) of 0.95, Intersection over Union (IoU) of 0.90, F1-score of 0.95, recall of 0.93, and precision of 0.97.
• The proposed model demonstrated a mean error of 0.41 mm in localizing needle shafts.

Conclusions:

• The study presents a novel, automated method for localizing and reconstructing prostate HDR-BT interstitial needles from 3D CT images.
• This computer-aided approach has the potential to significantly improve the detection and delineation of multi-catheters in clinical settings, thereby enhancing treatment quality.