Content-based image retrieval assists radiologists in diagnosing eye and orbital mass lesions in MRI

Background: Diagnoses of eye and orbit pathologies by radiological imaging is challenging due to their low prevalence and the relative high number of possible pathologies and variability in presentation, thus requiring substantial domain-specific experience. Purpose: This study investigates whether a content-based image retrieval (CBIR) tool paired with a curated database of orbital MRI cases with verified diagnoses can enhance diagnostic accuracy and reduce reading time for radiologists across different experience levels. Material and Methods: We tested these two hypotheses in a multi-reader, multi-case study, with 36 readers and 48 retrospective eye and orbit MRI cases. We asked each reader to diagnose eight orbital MRI cases, four while having only status quo reference tools available (e.g. Radiopaedia.org, StatDx, etc.), and four while having a CBIR reference tool additionally available. Then, we analyzed and compared the results with linear mixed effects models, controlling for the cases and participants. Results: Overall, we found a strong positive effect on diagnostic accuracy when using the CBIR tool only as compared to using status quo tools only (status quo only 55.88%, CBIR only 70.59%, 26.32% relative improvement, p=.03, odds ratio=2.07), and an even stronger effect when using the CBIR tool in conjunction with status quo tools (status quo only 55.88%, CBIR + status quo 83.33%, 49% relative improvement, p=.02, odds ratio=3.65). Reading time in seconds (s) decreased when using only the CBIR tool (status quo only 334s, CBIR only 236s, 29% decrease, p<.001), but increased when used in conjunction with status quo tools (status quo only 334s, CBIR + status quo 396s, 19% increase, p<.001). Conclusion: We found significant positive effects on diagnostic accuracy and mixed effects on reading times when using the CBIR reference tool, indicating the potential benefits when using CBIR reference tools in diagnosing eye and orbit mass lesions by radiological imaging.