Convolutional neural networks for improving image quality with noisy PET data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Positron Emission Tomography (PET) imaging is inherently noisy due to data sparsity.
• Machine learning techniques are increasingly explored to enhance PET image quality.
• Improving image quality is crucial for accurate clinical diagnosis.

Purpose of the Study:

• To investigate a deep learning approach for denoising PET images using a 3D convolution neural network (CNN).
• To evaluate the effectiveness of CNN-based denoising compared to conventional methods.
• To assess the impact of denoising on physician performance in lesion detection tasks.

Main Methods:

• A 3D CNN was trained to denoise reconstructed PET images from chest cancer patient data with emulated noise levels.
• Ground truth was established using full-count PET reconstructions.
• Benefits over Gaussian smoothing were quantified through physician-led image ranking and lesion detection tasks.

Main Results:

• CNN-denoised images were ranked equal to or better than Gaussian-smoothed images across all noise levels.
• Significant improvements in lesion contrast recovery and detectability were observed, particularly in low-count datasets.
• At 1 million true counts, CNN-denoised images showed a higher true positive detection rate (40%) compared to smoothed images (30%).

Conclusions:

• 3D CNN denoising offers significant improvements for very noisy PET images and moderate benefits for all noise levels.
• The technique showed limited benefit for detection performance at routine clinical count levels.
• Further research may optimize CNNs for clinical PET image enhancement.