Kaiwen Deng

AI-assisted reading interfaces often present a few passages from a known scientific document. We evaluated how much annotated evidence remains visible under fixed selection budgets using SciFact and QASPER. Without fine-tuning, we compared leading passages, lexical overlap, document-local BM25, MiniLM, and BGE; SPECTER provided an additional exploratory scientific-text comparison. On 338 SciFact development evidence sets, BGE top-4 sentence recall was 0.788 versus 0.663 for BM25 (paired difference 0.125; 95% claim-cluster bootstrap interval [0.078, 0.171]), and complete coverage was 77.2% versus 64.2%. Only 12 of 25 multi-sentence sets were completely covered (48.0%). In an internal document-disjoint subset, BGE recall gains under 96-, 128-, and 160-word caps ranged from 0.086 to 0.097. QASPER results include an original 89-question yes/no comparison and sensitivities across 883 answerable questions and 714 uniquely text-mapped questions. SPECTER showed mixed results across the two datasets. These retrospective analyses concern evidence visibility within known documents. The illustrative prototype returns selected text and source positions; neither the results nor the prototype establishes improved reader understanding or decisions.

Fundus images from the two eyes of one patient can share disease and acquisition cues, making matching performance difficult to interpret. We developed a controlled evaluation protocol for same-patient cross-eye matching that combines disease matching, image quality controls, and patient-level bootstrap resampling. Frozen features from an ImageNet-pretrained 50-layer residual network (ResNet-50) were evaluated in 300 patients from DeepDRiD, a retinal dataset with two photographs per eye and patient-disjoint splits. In a post hoc 34-patient test subset matched on ordered bilateral diabetic retinopathy grades and right-eye quality tier, the area under the receiver operating characteristic curve (AUROC) was 0.861 (interval envelope, 0.681–0.995). A separate 33-patient four-view analysis showed higher cross-eye than matched-negative similarity; the interval for the same-eye–cross-eye difference included zero. External validation in 244 patients from the independent PAPILA glaucoma dataset yielded AUROC 0.859 under bilateral diagnosis matching (interval envelope, 0.804–0.907). A learned hierarchical head achieved a higher validation AUROC point estimate than frozen features in DeepDRiD (0.882 versus 0.843) but a lower estimate in a previously inspected 137-patient fundus–optical coherence tomography cohort (0.849 versus 0.902). These results support cross-eye correspondence under the tested controls and show that source-cohort optimization does not ensure better transfer. The protocol provides a reproducible assessment of bilateral fundus representations across cohorts.