Investigating the limited performance of a deep-learning-based SPECT denoising approach: An observer-study-based...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Healthcare
• Image Processing

Background:

• Deep learning (DL) based denoising methods show limited performance in objective assessment of image-quality (OAIQ)-based studies for signal detection tasks.
• Investigating the reasons for this limited performance is crucial for advancing medical imaging techniques.

Purpose of the Study:

• To conduct a task-based characterization of a DL-based denoising approach for individual signal properties in single photon-emission computed tomography (SPECT) images.
• To evaluate the performance of a convolutional neural network (CNN)-based denoiser on low-count SPECT images across different signal sizes and signal-to-background ratios (SBRs).

Main Methods:

• A CNN-based denoiser was trained on simulated SPECT images with varying signal properties and noise levels.
• Signal-detection tasks were designed using a signal-known-exactly/background-known-statistically (SKE/BKS) framework.
• Performance was evaluated using an anthropomorphic channelized Hotelling observer (CHO) for five signal sizes and four SBRs.

Main Results:

• The DL-based denoising method did not improve signal-detection task performance, consistent with previous OAIQ studies.
• Observer-study-based characterization demonstrated no performance enhancement for any signal type.
• The study provided insights into DL denoising performance as a function of signal size and contrast.

Conclusions:

• DL-based denoising approaches require objective, task-based evaluation to understand their limitations.
• Observer-study-based characterization offers a mechanism to assess DL method sensitivity to specific object properties.
• This work emphasizes the need for rigorous evaluation of AI in medical imaging.