CT-free quantitative SPECT for automatic evaluation of %thyroid uptake based on deep-learning.

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Artificial Intelligence in Healthcare

Background:

• Quantitative thyroid single-photon emission computed tomography/computed tomography (SPECT/CT) traditionally relies on CT for attenuation correction and thyroid segmentation.
• Manual segmentation of the thyroid on CT is time-consuming and can introduce variability in %thyroid uptake measurements.
• Developing a CT-free approach for quantitative SPECT could streamline the imaging process and improve efficiency.

Purpose of the Study:

• To develop a deep-learning-based method for CT-free quantitative thyroid SPECT.
• To generate synthetic attenuation maps (μ-maps) using deep learning.
• To achieve automatic segmentation of the thyroid gland for accurate %thyroid uptake quantification.

Main Methods:

• Retrospective analysis of 650 quantitative thyroid SPECT/CT datasets.
• Utilized 3D U-Nets for synthetic μ-map generation from SPECT data and for automatic thyroid segmentation.
• Validated the generated μ-maps and segmentations against CT-derived ground truth and manual segmentations, respectively, using independent datasets.

Main Results:

• Synthetic μ-maps showed high correlation (R²=0.972) and minimal error compared to CT-derived μ-maps.
• Automatic thyroid segmentation achieved excellent results with a Dice similarity coefficient of 0.767 and minimal volume difference.
• %Thyroid uptake measurements using the CT-free method were comparable to the conventional SPECT/CT approach (p=0.1090).

Conclusions:

• Deep learning can effectively generate synthetic attenuation maps and perform automatic thyroid segmentation for quantitative SPECT.
• CT-free quantitative thyroid SPECT is feasible, enabling accurate %thyroid uptake evaluation.
• This AI-driven approach has the potential to simplify quantitative SPECT imaging workflows.