Report on the AAPM deep-learning spectral CT Grand Challenge

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• The 2022 AAPM Grand Challenge focused on Deep-Learning spectral Computed Tomography (DL-spectral CT) image reconstruction.
• Spectral CT offers advanced material decomposition capabilities beyond conventional CT.

Purpose of the Study:

• To develop highly accurate image reconstruction algorithms for dual-energy CT scans.
• To solve the inverse problem for three-tissue map decomposition using fast kilovolt switching.
• To compare deep learning, iterative, and hybrid reconstruction approaches.

Main Methods:

• A 2D breast CT simulation with stochastic tissue maps (adipose, fibroglandular, calcification) was used.
• Dual-energy scans simulated with alternating 50 and 80 kilovolts (kV) x-ray potentials.
• Deep learning models trained on 1000 cases, predicting material maps from transmission data or FBP images.

Main Results:

• 18 research groups submitted algorithms, with 17 utilizing deep learning (DL).
• The winning and second-place algorithms achieved near-zero root-mean-square error (RMSE) accuracy.
• Top-performing DL algorithms demonstrated exceptional accuracy in spectral CT image reconstruction.

Conclusions:

• The DL-spectral CT challenge successfully fostered innovation in image reconstruction for spectral CT.
• Deep learning approaches proved highly effective for solving the spectral CT inverse problem.
• The challenge established a benchmark for advanced spectral CT image reconstruction techniques.