An Interpretable Handcrafted-Feature Framework for Multi-Modal Image Quality Assessment of Gastric CT and Endoscopic Images

Abstract

The proposed IQA method provides a comprehensive, interpretable image quality assessment (IQA) framework that can be used for a comparative analysis of CT and endoscopic gastric images. This IQA includes handcrafted features such as LBP/ GLCM textures and clarity (Laplacian/Sobel/Tenengrad/Brenner), resolution (FFT-based), and morphological top-hat/black-hat abnormality scores. These features are combined into a multi-stage confidence score.
When applied to a dataset of 200 images (100 CT, 100 endoscopic) taken from publicly available sources, the Mann-Whitney U test with Holm-Bonferroni correction revealed that CT images have significantly higher clarity scores (0.788 ± 0.108 vs. 0.663 ± 0.169; p < .001, d = 0.88). Conversely, endoscopic images show significantly higher quality of textures (0.241 ± 0.027 vs. 0.159 ± 0.048; p < .001, d = −2.10) and significantly higher abnormality scores (0.576 ± 0.066 vs. 0.376 ± 0.023; p < .001, d = −4.06). The categorical distribution of the quality across different modalities was found to differ significantly (χ² = 15.41, p < .001), while the differences between the continuous overall confidence score and the resolution were not significant (p = .093 and p = .066, respectively). All null hypotheses are presented here together with corresponding post-hoc power analyses. The overall clarity is significantly and positively correlated with the overall confidence score (r = 0.937, p < .001).
Robustness analysis shows that the proposed method maintains its ranking stability in case of ±50% perturbation of input weights (Spearman ρ > 0.92). On the other hand, a moderate multicollinearity is detected for clarity and resolution inputs (VIF > 6). The method is explicitly proposed as a tool for characterizing image modalities and not as a replacement for any supervised methods.