MaskRet-Oto: Region-Conditioned Evidence Retrieval and Calibration for Otoscopic Image Classification and Selective Referral

Abstract Digital otoscopy is a compact biomedical imaging modality, but retrospective image-classification results can be distorted by duplicate content, acquisition-domain cues, uncertain confidence, and dataset prevalence. We revised the framework as MaskRet-Oto, a region-conditioned evidence-retrieval and calibration pipeline rather than a retrieval-augmented generation (RAG) system. The working archive contained 3,000 RGB otoscopic images in five approximately balanced categories. Exact-hash auditing identified 78 duplicated files; the remaining 2,922 images formed 1,956 within-class perceptual-hash groups that were kept intact across training (n=2,046), validation (n=438), and test (n=438) partitions. Because patient identifiers were unavailable, this procedure controls only exact/near-duplicate image leakage and cannot establish patient-independent evaluation; different views from the same patient or examination could remain in different splits. The soft central field-of-view operator retained 70.9% of image support and is now described as a predefined region-conditioning prior, not as empirical proof that shortcut learning was prevented. On the internal image-level test set, the compact 64 x 64 CNN pipeline achieved 98.17% accuracy (95% bootstrap confidence interval 96.80-99.32%), 98.23% macro-F1, MCC 0.977, and macro-AUC 0.9994. Training-only visual retrieval reached 99.54% top-5 recall. The source-tagged textual memory contains 15 curated statements retrieved by TF-IDF; no language model is used and the component is not presented as a clinical knowledge base or RAG. Retrieval fusion did not improve top-1 accuracy; its role is limited to exposing similar training cases, attaching source-tagged evidence, and contributing a small additional calibration change after temperature scaling. Selective prediction concentrated all observed errors in the lowest-confidence deciles in this same test set, but this is a post hoc internal utility surrogate, not a clinical-utility validation. The nearly balanced benchmark distribution also means that calibration values should not be interpreted as real-world disease-risk probabilities. Patient-level, device-diverse, prevalence-aware, and prospective validation with stronger standard backbones is required before clinical use.