Author: Hilary P Bagshaw, Mark K Buyyounouski, Serdar Charyyev, Xianjin Dai, PhD, Yu Gao, Thomas R. Niedermayr, Lei Xing 👨🔬
Affiliation: Department of Radiation Oncology, Stanford University 🌍
Purpose: Real-time transrectal ultrasound imaging is the gold standard for needle placement and treatment planning of real-time based-ultrasound-based high dose-rate (HDR) prostate brachytherapy. Cumulative needle artifact in the ultrasound image occurs that presents a challenge interpreting needle placement and contouring treatment volumes, particularly for learners and novices. The aim of this study is to enhance the quality of ultrasound imaging by erasing needles and their associated artifacts using artificial intelligence (AI).
Methods: Dataset from 100 patients with prostate cancer treated with HDR brachytherapy were retrospectively collected. For each patient, a set of ultrasound images immediately before needle insertion and another set immediately after needle insertion were acquired using the bkSpecto Ultrasound System (BK Medical, Herlev, Denmark). A Cycle Generative Adversarial Network (cyclyGAN) model was trained to enhance the image quality and erase needles and associated artifacts from the post-needle scan. Images from 80 patients (5289 slices) were used for model training, and images from the remaining 20 patients (1164 slices) were used for testing. The model was trained with 200 epochs using a NVIDIA GeForce RTX™ 4090 GPU.
Results: It took the trained network 192 seconds to generate needle-free images for the 20 testing patients (<10s per patient). The trained network is capable of enhancing image quality and eliminating the appearance of needles and associated artifacts. Following the AI-based enhancement, the synthetic images demonstrate a greater degree of similarity to the original pre-needle volumes, and the prostate and urethra can be easily identified.
Conclusion: A deep learning-based network was trained to erase needles and their associated artifacts in the post-needle ultrasound images. This enhancement eases needle insertion and contouring. As a result, this AI-based needle eraser has the potential to reduce procedure time, minimize errors, and facilitate the training of brachytherapy specialists.