Approximating the Hotelling observer with autoencoder-learned efficient channels for binary signal detection tasks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Quality Assessment

Background:

• Objective assessment of medical image quality (IQ) is crucial for system optimization.
• The Hotelling observer (HO) is an optimal linear observer for IQ metrics, but conventional methods are often computationally intractable.
• Channelized methods approximate the HO but vary in performance depending on imaging conditions and tasks.

Purpose of the Study:

• To present and implement a channelized HO method using an autoencoder (AE) to characterize its performance.
• To investigate the efficiency of AE-learned channels for approximating the HO.
• To evaluate the performance of the proposed AE-based method against state-of-the-art channelized methods.

Main Methods:

• Training an AE was shown to be equivalent to developing channels for HO approximation.
• The AE loss function was modified to incorporate task-relevant information, increasing channel efficiency.
• Performance was evaluated using binary detection tasks with phantom backgrounds and varying dataset sizes, including generalization studies.

Main Results:

• AE-learned channels demonstrated comparable performance to state-of-the-art methods in detection tasks.
• Superior performance was observed in generalization studies for AE-learned channels.
• Incorporating a signal prior significantly improved performance, especially in small datasets.

Conclusions:

• Autoencoders (AEs) can learn efficient channels for the Hotelling observer (HO).
• The proposed method shows promising implications for future medical imaging technology assessments.
• Improved detection performance with small datasets using a signal prior highlights the method's potential.