Transfer learning‑based attenuation correction in ^99mTc-TRODAT-1 SPECT for Parkinson's disease using realistic...

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence

Background:

• Dopamine transporter (DAT) SPECT is crucial for early Parkinson's disease (PD) detection.
• Image attenuation significantly hampers DAT SPECT accuracy.
• Deep learning (DL) methods show promise for attenuation correction (AC).

Purpose of the Study:

• To investigate the efficacy of transfer learning (TL) using realistic Monte Carlo (MC) simulation data for deep learning-based attenuation correction (DLAC) in DAT SPECT.
• To enhance AC performance in DAT SPECT by leveraging pre-trained models for TL.

Main Methods:

• A 3D conditional generative adversarial network (cGAN) was pre-trained on simulated SPECT data (NAC/CTAC).
• The pre-trained model was fine-tuned using limited clinical DAT SPECT data (TLAC-MC).
• Performance was compared against various DLAC methods and Chang's AC.

Main Results:

• TLAC-MC achieved superior performance in NMSE and SSIM compared to other methods, particularly with limited clinical data (8 datasets).
• Performance improvements were observed as the number of fine-tuning datasets increased.
• TLAC demonstrated greater benefits when the pre-training data domain closely matched the target domain.

Conclusions:

• Transfer learning attenuation correction (TLAC) is feasible for DAT SPECT using simulation-based pre-trained models.
• TLAC-MC offers superior AC performance, especially in low-data clinical scenarios.
• The domain similarity between pre-training and target data is critical for effective TLAC.