Comparative Evaluation of Machine Learning and Conventional Material Decomposition Algorithms for Spectral Chest...

Area of Science:

• Medical Imaging Physics
• Computational Imaging
• Radiographic Detector Technology

Background:

• Spectral chest radiography using photon-counting detectors (PCDs) allows for energy-resolved imaging, crucial for differentiating bone and soft tissues.
• The quantitative accuracy of PCD-based spectral imaging is significantly influenced by detector cross-talk and the choice of material decomposition algorithms.

Purpose of the Study:

• To compare the quantitative performance and low-contrast detectability of a conventional polynomial decomposition model against two machine learning (ML) algorithms (MLP and SVR) for cadmium telluride (CdTe) PCD spectral chest radiography.
• To evaluate the impact of varying calibration grid densities, energy thresholds, and radiation dose levels on the performance of these decomposition methods.

Main Methods:

• A physics-based simulation incorporating a Geant4-derived detector response model, charge transport, charge sharing, and Poisson noise was developed.
• Quantitative bias and RMSE were assessed using digital phantoms with known material compositions (aluminum, PMMA).
• Task-based low-contrast detectability was evaluated using the exponential transformation of the free-response operating characteristic (EFROC) method.

Main Results:

• Polynomial decomposition demonstrated stability with sparse calibration, while ML methods (MLP, SVR) significantly benefited from denser calibration grids and increased energy thresholds.
• Support vector regression (SVR) yielded the lowest root mean square error (RMSE) under dense calibration conditions.
• Multilayer perceptron (MLP) produced smoother image maps and enhanced soft-tissue detectability at low-to-intermediate radiation doses.

Conclusions:

• Machine learning algorithms, especially SVR and MLP, offer superior quantitative accuracy and low-contrast detectability in PCD spectral chest radiography compared to traditional polynomial methods, particularly with optimized calibration and energy threshold settings.
• Practical trade-offs exist between the calibration data requirements, achievable quantitative precision, and the detectability of subtle soft-tissue structures when implementing ML-based decomposition algorithms for PCD spectral imaging.