Fig. 7.

Noise invariance testing of deep learning networks across varying noise levels and frame averaging - Noise-level invariance was evaluated using three phantom datasets: (a)tree branches, (b) lead pieces, and (c) metal screws. (d, e, f) Across all frame averaging cases, deep learning networks (UN-4 layer, R2-UN, and Dense-UN) demonstrated significantly higher PSNR values than their corresponding inputs. For 32-frame averaging, the residual noise in outputs led to non-significant SSIM improvements for UN-4 layer and R2-UN, but Dense-UN achieved significantly higher SSIM values. As frame averaging increased, image quality improvements plateaued, and the networks converged to similar performance observed in (f).