Material decomposition from photon-counting CT using a convolutional neural network and energy-integrating CT...

Area of Science:

• Medical Imaging
• Computational Imaging
• Radiology

Background:

• Photon-counting CT (PCCT) offers superior dose efficiency and spectral resolution compared to energy-integrating CT (EID), crucial for material decomposition.
• Spectral distortions in the photon-counting detector (PCD) currently limit the accuracy of PCCT-based material decomposition.
• Accurate material decomposition is vital for advanced applications like theranostics.

Purpose of the Study:

• To develop and evaluate a deep learning (DL) approach to compensate for PCD spectral distortions.
• To improve the accuracy of material decomposition in PCCT using DL.
• To leverage high-dose multi-EID data for training DL models for PCCT material decomposition.

Main Methods:

• A 3D U-net deep learning architecture was employed for spectral distortion compensation.
• Training labels were derived from decomposition maps generated by high-dose multi-EID data.
• Performance was compared using PCD filtered back projection (FBP), iterative reconstruction (Iter), and direct decomposition (Decomp) as inputs.

Main Results:

• The DL approach, particularly with iterative reconstruction input (Iter2Decomp), significantly outperformed PCD matrix inversion decomposition.
• Iter2Decomp reduced RMSE by 27.50% for iodine and 59.87% for photoelectric effect maps, while increasing SSIM.
• While effective, the DL approach introduced some blurring, reducing the modulation transfer function (MTF) from 1.98 to 1.75 line pairs/mm at 50% MTF.

Conclusions:

• The demonstrated DL approach effectively generates more accurate material maps from PCCT data by correcting spectral distortions.
• This method shows excellent agreement with multi-EID decomposition, validating its accuracy for preclinical applications.
• Improved spectral PCCT imaging holds potential for advancing nanoparticle development in theranostics.