Comparative Evaluation of Machine Learning and Conventional Material Decomposition Algorithms for Spectral Chest Radiography Using a CdTe Photon-Counting Detector

Spectral chest radiography with photon-counting detectors (PCDs) enables energy-resolved acquisition for bone/soft-tissue separation, but quantitative performance depends on detector cross-talk and the selected material decomposition algorithm. We performed a controlled simulation study comparing a conventional low-order polynomial decomposition model with two machine learning regressors (multilayer perceptron (MLP) and support vector regression (SVR)) for a cadmium telluride (CdTe) PCD. A Geant4-derived detector response model, coupled with a charge-transport model, was integrated into a physics-forward model including charge sharing and Poisson quantum noise. Digital LucAl/IEC 62220-2-1 phantoms with aluminum and polymethyl methacrylate inserts were used for quantitative bias/root mean square error (RMSE) evaluation, and task-based low-contrast detectability that was evaluated using an exponential transformation of the free-response operating characteristic (EFROC) method using a matched-filter template. Performance was evaluated over clinically relevant dose levels (0.07–7.5 mAs), calibration grid densities (3×3 to 8×8), and numbers of energy thresholds (M=2–6). Polynomial decomposition was stable under sparse calibration, whereas ML methods benefited strongly from denser calibration and additional thresholds; SVR achieved the lowest RMSE under dense calibration, while MLP produced smoother maps and improved soft-tissue detectability at low-to-intermediate dose. At high dose, all methods approached near-ideal detection performance. These results quantify practical trade-offs between calibration requirements, quantitative accuracy, and low-contrast detectability for PCD-based spectral chest radiography.