Photon-counting CT thermometry via material decomposition and machine learning

TL;DR

This paper introduces a novel photon-counting CT method combined with machine learning to improve real-time temperature measurement during thermal ablation, demonstrating promising accuracy with experimental data.

Contribution

It develops a new approach using spectral photon-counting CT and neural networks for accurate, nonlinear temperature estimation in thermal therapies.

Findings

Neural network achieved mean absolute error of 1.80°C on 300 mM CaCl2.

Method successfully predicted temperature with low error on different materials.

Feasibility demonstrated with experimental spectral data and neural network modeling.

Abstract

Thermal ablation procedures, such as high intensity focused ultrasound (HIFU) and Radiofrequency Ablation (RFA), are often used to eliminate tumors by minimally invasively heating a focal region. For this task, real-time 3D temperature visualization is key to target the diseased tissues while minimizing damage to the surroundings. Current CT thermometry is based on energy-integrated CT, tissue-specific experimental data, and linear relationships between attenuation and temperature. In this letter, we develop a novel approach using photon-counting CT for material decomposition and a neural network to predict temperature based on thermal characteristics of base materials and spectral tomographic measurements of a volume of interest. In our feasibility study, distilled water, 50 mM CaCl2, and 600 mM CaCl2 are chosen as the base materials. Their attenuations are measured in four discrete energy bins at various temperatures. The neural network trained on the experimental data achieves a mean absolute error of 1.80 {\deg}C and 3.97 {\deg}C on 300 mM CaCl2 and a milk-based protein shake respectively. These experimental results indicate that our approach is promising for handling nonlinear thermal properties for materials that are similar or dissimilar to our base materials.