Muhammet Fatih Demir

Background/Objectives: Accurate diagnosis of middle ear disease remains challenging because otoscopic interpretation is subjective, while many artificial intelligence models depend on manually selected still frames from video examinations. We aimed to develop and internally evaluate LUCID, a systematic framework for automatically identifying the most informative frames in otoscopy videos. Methods: We analyzed 713 otoscopy videos from 491 patients. LUCID combines a ResNet-50 classifier for coarse eardrum-visibility assessment, BC-AdvCAM for weakly supervised eardrum localization and coverage estimation, and an otoscope-specific blur/focus assessment into a within-video frame-ranking score. Automated selections were compared with expert-selected frames through blinded human review and downstream classification. To separate ranking quality from the effects of bag size and multi-frame aggregation, a matched 12-frame control using uniform temporal sampling was evaluated with the same ResNet-50 and attention-based multiple instance learning (ABMIL) pipeline. Results: Among 293 video pairs with complete ratings, algorithm-selected frames were rated equal to or better than expert-selected frames in 74.7% of evaluations on average. However, when an evaluator expressed a non-Equal preference, the expert frame was favored in 71.7% and 70.7% of cases by the two evaluators, respectively; among concordant non-Equal ratings, 80.0% favored the expert frame. BC-AdvCAM achieved an intersection over union of 0.6717 and a Dice similarity coefficient of 0.7924 on a held-out manually segmented test subset. For downstream diagnosis, patient-level mean accuracy for the matched single-frame comparison was 68.79% for LUCID Top-1 versus 72.17% for the expert-selected frame (difference, −3.38 percentage points; 95% exact sign-flip-inversion CI, −6.79 to 0.00; two-sided exact patient sign-flip p = 0.0540). In the matched 12-frame comparison, LUCID Top-12 achieved 78.13% versus 59.41% for Uniform-12 using the identical ABMIL pipeline (difference, +18.71 percentage points; 95% exact sign-flip-inversion CI, +15.79 to +21.63; two-sided exact patient sign-flip p = 2.33 × 10⁻¹⁰). Conclusions: In the matched single-frame comparison, the expert-selected frame had the higher observed patient-level mean accuracy, whereas LUCID provided an automated ranking of the full video that substantially improved matched 12-frame ABMIL performance relative to uniform temporal sampling. These results support LUCID as an automated frame-ranking and multi-frame evidence-selection tool, while external validation remains necessary before clinical deployment.

Accurate diagnosis of middle ear diseases using otoscopy remains challenging, particularly in primary care settings where clinician experience with otoscopy varies and visual examination alone may not fully capture middle ear physiology. Although AI-based models have shown promise for automated otoscopy interpretation, most rely primarily on visual information, creating an opportunity to improve diagnostic robustness by integrating complementary physiological measurements. Here, we present OtoTymp-AI, a confidence-guided multimodal decision-fusion framework that integrates otoscopy video analysis with conventional 226 Hz tympanometry. A convolutional neural network was trained to classify otoscopy videos, and tympanometric measurements were incorporated through a clinically interpretable, rule-guided decision strategy to refine visually uncertain predictions rather than through jointly trained multimodal representation learning. Within this multi-center otoscopy study, multimodal evaluation was performed in one independent external cohort comprising 104 videos with paired tympanometry across six diagnostic categories. In this exploratory paired-cohort analysis, the video-only CNN achieved an overall accuracy of 63.46% (95% CI, 53.88–72.08), while the hybrid video-plus-tympanometry framework achieved 81.73% accuracy (95% CI, 73.22–87.98) at the best observed threshold of 0.90, corresponding to an observed absolute improvement of 18.27 percentage points over the CNN-only model (95% paired bootstrap CI, 9.62–26.92). Improvements were observed for selected categories represented in this cohort, including effusion and retraction, although confidence intervals were wide for categories with small positive sample sizes. These findings suggest that integrating anatomical video information with physiological tympanometry may improve AI-assisted middle ear diagnosis in this external paired cohort, but the operating threshold and per-class performance require confirmation in larger prospective paired multimodal cohorts.

Background/Objectives: Digital pathology models increasingly infer molecular phenotypes from hematoxylin and eosin whole-slide images (WSIs), but most assume that one slide adequately represents patient-level tumor biology. We quantified prediction instability caused by slide selection and evaluated multi-WSI aggregation. Methods: Six multiple-instance learning architectures were developed in task-specific subsets of TCGA-BRCA (source cohort: 1006 patients and 1065 WSIs) and independently evaluated in 122 patients (400 WSIs) from CPTAC-BRCA. The clinically realistic comparison was random one-slide-per-patient inference versus patient-level aggregation of all available WSIs. Label-conditioned best- and worst-slide analyses were used only as retrospective oracle bounds. The recurrence-risk endpoint was a research-derived 21-gene recurrence-score surrogate calculated from RNA sequencing rather than a clinically reported Oncotype DX result. Results: Random single-slide selection yielded AUCs of 0.77–0.86 for recurrence-risk prediction and 0.65–0.77 for HER2 status. Patient-level multi-WSI aggregation yielded AUCs of 0.81–0.89 and 0.71–0.80, respectively. The architecture-controlled AUC improvement from random selection to aggregation ranged from 0.026 to 0.041 for recurrence risk and from 0.036 to 0.073 for HER2. Aggregation also reduced Brier scores by 0.010–0.028 and 0.009–0.106, respectively. The label-conditioned oracle analyses demonstrated wider theoretical performance ranges of 0.46–0.97 and 0.40–0.92, respectively. Slide-selection sensitivity persisted after excluding tumor-evidence-negative and low-tumor-content WSIs. Conclusions: Slide selection is a material source of sampling-dependent prediction uncertainty. Patient-level multi-WSI aggregation mitigated selection-dependent degradation across architectures, although validation using routine institutional material and clinically reported molecular assays remains necessary.