Abstract
Multimodal medical image super-resolution remains challenging, as generative models must increase resolution while preserving anatomical integrity and avoiding reconstruction biases. For robust model assessment, we generate realistic low-resolution PET and MRI inputs using k-space degradation based on Fourier transformation, ensuring structural detail is retained for fair evaluation. Using these LR inputs, we compare SRAE and SRGAN within a multitask PET/MRI super-resolution framework and assess trustworthiness with a Vision Transformer and Extreme Value Theory (ViT-EVT), which assigns pixel-level anomaly classes using statistically defined thresholds. In addition, we benchmark our approach against state-of-the-art baselines (PatchCore and quantile-based detection), showing that ViT-EVT provides the most accurate anomaly classification across all classes and enables a clearer interpretation of reconstruction behaviour. Outcomes show that SRGAN produces sharper and more accurate reconstructions than SRAE, improving PET performance in the worst case from SRAE, given by R2=0.60, to 0.80 and MRI performance to 0.90, while reducing high-anomaly pixels from 14.4 to 9.7% and strong anomalies from 0.6 to 0.3%. Using the confusion-matrix evaluation, ViT-EVT reaches Normal-class accuracies above 91% and retains only 7–30% of High and Strong deviations, quantifying its ability to distinguish regions reconstructed with high fidelity from those where anomalies persist. The framework demonstrates meaningful gains in reconstruction fidelity and reliability.