Background: Accurate reconstruction of internal temperature fields from surface temperature data is critical for applications such as non-invasive thermal imaging, particularly in scenarios involving small temperature gradients, like those in the human body. Methods: In this study, we employed 3D convolutional neural networks (CNNs) to predict internal temperature fields. The network’s performance was evaluated under both ideal and non-ideal conditions, incorporating noise and background temperature variations. A physics-informed loss function embedding the heat equation was used in conjunction with statistical uncertainty during training to simulate realistic scenarios. Results: The CNN achieved high accuracy for small phantoms (e.g., 10 cm in diameter). However, under non-ideal conditions, the network’s predictive capacity diminished in larger domains, particularly in regions distant from the surface. The introduction of physical constraints in the training processes improved the model’s robustness in noisy environments, enabling accurate reconstruction of hot-spots in deeper regions where traditional CNNs struggled. Conclusions: Combining deep learning with physical constraints provides a robust framework for non-invasive thermal imaging and other applications requiring high-precision temperature field reconstruction, particularly under non-ideal conditions.

Background/Objectives: Modern tomographic reconstruction methods—including physics-informed and AI-based approaches—can produce very high fidelity images. In this regime, widely used global image quality metrics often approach saturation, making it harder to distinguish residual differences between methods and identify remaining performance gaps. This study introduces a physically grounded and standardized evaluation framework designed to retain sensitivity beyond conventional global metrics and support both comparison and systematic improvement in tomographic reconstruction methods. Methods: The proposed framework defines standardized reference images—“Source”, “Detector”, “Ideal”, and “Realistic”—using Monte Carlo simulations, with the Ideal Image serving as a physically grounded benchmark. Reconstruction performance is evaluated using pixel-wise difference and ${\chi }^2$ maps, Region-of-Interest analysis, intensity (gray-value) histogram comparisons, and the Structure and Contrast Index (SCI), computed on difference maps. Demonstrations use simulated SPECT data reconstructed with ART, MLEM, and RISE-1. Results: Across case studies, SCI and ${\chi }^2$-based diagnostics reveal structured residuals and localized deficiencies not evident from global similarity metrics such as SSIM or NMSE. Comparative analyses show that methods with similar global scores can exhibit distinct residual structures and region-specific performance variations, while improved agreement in the sinogram domain does not necessarily translate into improved image fidelity. Histogram-based diagnostics provide complementary information on intensity redistribution not captured by pixel-domain summaries. Conclusions: The framework provides a reproducible, physically meaningful, and sensitive approach for evaluating tomographic reconstruction performance in the high-fidelity regime. By combining standardized reference images with multi-domain and multi-metric analysis, it enables robust benchmarking and supports physically consistent interpretation of reconstruction quality.

In this study, we explore the use of 3D convolutional neural networks (CNNs) for predicting internal temperature fields from the surface temperature, with a focus on applications where small temperature gradients, similar to those in the human body, are present. The network accuracy was evaluated under both ideal and non-ideal conditions which include noise and background temperature effects. In non-ideal scenarios, the network accurately reconstructed the 3D temperature field for small phantoms (e.g., 10 cm in diameter). However, as the size of the domain increased, the network’s predictive capacity diminished, particularly in regions far from the surface. To address this limitation, we introduced statistical uncertainty during training, simulating non-ideal conditions, in combination with a physics-informed loss function which embed the heat equation directly into the training process. This combination can improve the model’s performance, particularly in noisy environments, where traditional CNN architectures failed to reconstruct hot-spots in deeper regions. Our results suggest that combining deep learning with physical constraints offers a robust framework for non-invasive thermal imaging and other applications requiring high-precision temperature field reconstruction.