Answers are grounded in this paper and cite their sources.
Background Otoscopic image screening can be compromised by duplicate leakage, class imbalance, acquisition shortcuts and uncalibrated confidence. We assess a duplicate-audited, three-representation fusion strategy with selective referral rather than multiview camera reconstruction. Methods A retrospective 699-image archive yielded 687 integrity-audited images spanning AOM, ASOM, CSOM and Normal categories. Three deterministic representations of each photograph (unmasked, geometric ellipse and brightness-refined) were encoded using handcrafted descriptors. Each representation was modeled by a four-class one-versus-rest logistic-regression system (four binary class-vs-rest estimators per representation), and the three normalized four-class posterior vectors were averaged equally. Original evaluation used 25 repeated duplicate-grouped fold estimates and a previously explored 115-image internal reference partition. Class-wise precision/recall/F1 and descriptive uncertainty were reconstructed from the reported confusion matrix. Results Reported grouped mean macro F1 was 0.674 ± 0.044 (macro AUC 0.875 ± 0.027). Internal-reference macro F1 was 0.735 (macro AUC 0.889), versus 0.704 for the strongest individual representation (descriptive difference +0.031). Internal-reference F1 was 0.579 AOM, 1.000 ASOM, 0.500 CSOM and 0.861 Normal. Reported temperature scaling lowered ECE from 0.324 to 0.123. A confidence threshold of 0.70 retained 60/115 images, correctly classified 53/60 (88.3%) and referred 16/23 observed errors; the retained accuracy missed the development target of 90%. Conclusions The reported findings demonstrate an exploratory, image-level analytical workflow, not patient-independent or externally validated clinical screening. ASOM was confounded with JPEG format; label adjudication, nested tuning, full hyperparameter recovery, patient-linked validation, matched ablations and group-wise calibration remain unresolved. Confidence or abstention alone is not an explanation or a medical safety guarantee.
No takes yet. Share an insight, caveat, or question.
Singh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: