Abstract
Diagnosing eye and orbit pathologies through radiological imaging presents considerable challenges due to their low prevalence, the extensive range of possible conditions, and their variable presentations, necessitating substantial domain-specific expertise. This study evaluates whether a ML-based content-based image retrieval (CBIR) tool, combined with a curated database of orbital MRI cases with verified diagnoses, can enhance diagnostic accuracy and reduce reading time for radiologists diagnosing eye and orbital pathologies. It explores whether this tool alone, or in combination with status quo reference tools (e.g. Radiopaedia.org, StatDx) provides these benefits. In a multi-reader, multi-case study involving 36 radiologists and 48 retrospective orbital MRI cases, participants diagnosed eight cases: four using status quo reference tools and four with the addition of the CBIR tool. Analysis using linear mixed-effects models revealed significant improvements in diagnostic accuracy when using the CBIR tool alone (55.88% vs. 70.59%, p = 0.03, odds ratio = 2.07) and an even greater improvement when used alongside status quo tools (55.88% vs. 83.33%, p = 0.02, odds ratio = 3.65). Reading time decreased when using the CBIR tool alone (334 s vs. 236 s, p < 0.001) but increased when used in conjunction with status quo tools (334 s vs. 396 s, p < 0.001). These findings indicate that CBIR tools can significantly enhance diagnostic accuracy for eye and orbit diagnostics, though their impact on reading time varies.