Microwave thermal therapy produces spatially localized heating whose precise numerical representation depends on energy deposition, thermal conduction, and blood perfusion. In a newly published comparative study, researchers Ahmad H. Milyani and Nidal H. Abu-Hamdeh investigated how accurately three compact coordinate-based regression families could reconstruct withheld portions of a single published steady-state finite-element bioheat field using sparse nodal samples.
The comparative research assessed three distinct regression architectures: ridge-regularized tensor-product B-splines, a Fourier-feature multilayer perceptron, and Chebyshev Elastic Net. To ensure a balanced and systematic comparison, the performance of each family was evaluated using matched twenty-evaluation optimization searches driven by differential evolution as well as dual-annealing algorithms.
To reduce immediate-neighbor data leakage across the domain, the experimental design incorporated five spatially blocked test folds, a separate validation fold, and 0.5-millimeter exclusion buffers around evaluation boundaries. The modeling framework examined performance across three independent sampling seeds and two distinct training allocations, testing sparse budgets of 649 nodes, which correspond to one percent of the nodes, and 3,243 nodes, corresponding to five percent of the nodes.
The ridge-regularized tensor-product B-spline achieved the lowest overall errors across the benchmarked tasks. In its best optimizer configurations, the spline attained a mean held-out mean absolute error of 0.701 Kelvin under the smaller 649-node budget and reduced the error to 0.450 Kelvin under the 3,243-node budget. At that larger five-percent training budget, the spline configuration also reached a 345-Kelvin node-mask Dice score of 0.899 within hot-bearing spatial folds.
Additional sensitivity analyses conducted with a wider exclusion buffer and an alternative spatial block layout similarly favored a spline configuration in the overall aggregate, even though specific performance rankings were dependent on the chosen block layout. Ultimately, these findings identify effective sparse-reconstruction representations and highlight regional error behavior across this solved thermal field, while explicitly avoiding any assumption of direct transfer to untested therapy configurations.
Subject of Research: Sparse reconstruction of finite element microwave bioheat fields
Article Title: Machine learning-based reconstruction of finite element thermal fields induced by microwave heating: a comparative study
Article References: Milyani, A. H., & Abu-Hamdeh, N. H. (2026). Machine learning-based reconstruction of finite element thermal fields induced by microwave heating: a comparative study. Scientific Reports, 16(1), Article 38925. https://doi.org/10.1038/s41598-026-81107-7
Image Credits: AI Generated
DOI: 10.1038/s41598-026-81107-7
Keywords: microwave thermal therapy, bioheat field, finite element reconstruction, sparse sampling, coordinate-based regression, tensor-product B-splines, Fourier-feature multilayer perceptron, Chebyshev Elastic Net, differential evolution, dual annealing, mean absolute error, Dice score, spatial blocking, exclusion buffers, temperature reconstruction, steady-state bioheat
Cite Scienmag News
Gavin Prescott. (November 22, 2026). Ridge-Regularized Splines Reconstruct Finite Element Microwave Heating Fields From Sparse Samples. Scienmag. https://scienmag.com/ridge-regularized-splines-reconstruct-finite-element-microwave-heating-fields-from-sparse-samples/
Gavin Prescott. "Ridge-Regularized Splines Reconstruct Finite Element Microwave Heating Fields From Sparse Samples." Scienmag, 22 November 2026, https://scienmag.com/ridge-regularized-splines-reconstruct-finite-element-microwave-heating-fields-from-sparse-samples/. Accessed 22 November 2026.
Gavin Prescott. "Ridge-Regularized Splines Reconstruct Finite Element Microwave Heating Fields From Sparse Samples." Scienmag. November 22, 2026. https://scienmag.com/ridge-regularized-splines-reconstruct-finite-element-microwave-heating-fields-from-sparse-samples/








