Fig. 8 | Scientific Reports

Fig. 8

From: Performance of a GPU- and time-efficient pseudo-3D network for magnetic resonance image super-resolution and motion artifact reduction

Fig. 8

Quantitative and qualitative evaluation of uncertainty. (a) GT image; (b) SR image; (c) SSIM map; (d) absolute error map between GT and SR images; (e) aleatoric uncertainty map rescaled to the range of 0 to 1 , illustrating its widespread presence across the image volume, particularly in the background where random noise dominates; (f) epistemic uncertainty map also rescaled to the range of 0 to 1, emphasizing anatomical structures: regions with high epistemic uncertainty correspond to lower SSIM values and higher absolute error values; (g) Linear regression for SSIM and (h) exponential regression for PSNR against epistemic uncertainty, using the accuracy test data (blue triangles) achieved \({R}^{2}\) values above 0.8, indicating a strong fit. Additionally, 93.7% and 95.5% of accuracy prediction data (green crosses) fell within the 95% predict intervals of SSIM and PSNR, respectively.

Back to article page