Abstract

Deep learning (DL) methods are currently being explored to recover images from sparse-view, limited-data, and undersampled acquisitions in medical applications. Although DLbased solutions may appear visually appealing based on likability/subjective criteria (such as less noise, smooth features), they may also suffer from imperceptible fakes. This issue is further exacerbated by a lack of easy-to-use techniques and robust metrics for the identification of fakes in DL-based outputs. In this work, we propose performing Fourier Ring Correlation (FRC)based analysis over small patches and concomitantly scanning across DL-based outputs and their reference counterparts to identify fakes. We term the metrics as sFRC. We describe the rationale behind sFRC and provide its mathematical framework. The thresholds required for the sFRC can be set using predefined fake features or imaging theory-based fake maps. We use sFRC to identify fakes for two undersampled medical imaging problems (CT super-resolution and MRI subsampled recovery). We demonstrate the effectiveness of sFRC in finding fake features for the two imaging problems and its agreement with a different imaging theory-based method on fake feature maps. Finally, we quantify the incidences of fakes from DL-based methods relative to indistribution versus out-of-distribution data and the increment in subsampling rate.