Benchmarking CNN, Transformer, Hybrid, and Foundation Models for Multi-Label Bronchoscopic Finding Classification on the BERD Dataset

Background: Automated classification of bronchoscopic findings is hindered by severe class imbalance, visual similarity between findings, and scarce expert-annotated data. This study benchmarked convolutional, Transformer, hybrid, and foundation models for multi-label bronchoscopic finding classification. Methods: ResNet-50, EfficientNet-B3, ViT-B/16, Swin-T, and Hybrid-EffSwin (average-pooling fusion of EfficientNet-B3 and Swin-T) were trained on 23 finding categories of the Bronchoscopy Examination Report Dataset (BERD) using identical hyperparameters and patient-level data splits. ConvNeXt-T and two foundation models (DINOv2 ViT-B/14 and BiomedCLIP ViT-B/16) served as additional baselines. Ablations examined fusion strategy, loss function, and decision thresholds. Models were compared using patient-level bootstrap confidence intervals (CIs) and Holm-corrected Wilcoxon tests on patient-level summaries. Results: On the independent test set (316 images, 180 patients), Hybrid-EffSwin achieved the highest macro F1-score (0.376), Matthews correlation coefficient (0.339), and area under the ROC curve (0.803). Across three training seeds, Hybrid-EffSwin ranked first in every run (macro F1 = 0.363 \((\pm)\) 0.013); its advantage over ResNet-50 and ViT-B/16 was significant in all runs, over EfficientNet-B3 in one of three, and over Swin-T in none (\((\Delta)\)F1 = +0.020 to +0.025).Among additional baselines, DINOv2 performed comparably to Hybrid-EffSwin (Hybrid-EffSwin minus DINOv2: \((\Delta)\)F1 = +0.013; 95% CI \((-)\)0.025 to 0.052) and achieved the best exact-match accuracy and calibration, whereas ConvNeXt-T was significantly inferior; a late-fusion ensemble of separately trained EfficientNet-B3 and Swin-T performed close to Hybrid-EffSwin. Average-pooling fusion performed at least as well as residual-projection and concatenation fusion, and validation-selected thresholds changed macro F1 only modestly. Conclusions: Hybrid CNN-Transformer, hierarchical Transformer, and self-supervised foundation models formed the strongest tier for multi-label bronchoscopic classification and could not be distinguished on the current test set. Multi-institutional validation is warranted.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11268987/v1
Primary Topic
AI in cancer detection
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Benchmarking CNN, Transformer, Hybrid, and Foundation Models for Multi-Label Bronchoscopic Finding Classification on the BERD Dataset

Ilkay Sibel Kervancı
Research Square
AI in cancer detection
preprint

Benchmarking CNN, Transformer, Hybrid, and Foundation Models for Multi-Label Bronchoscopic Finding Classification on the BERD Dataset

Ilkay Sibel Kervancı
preprint en

Abstract

Background: Automated classification of bronchoscopic findings is hindered by severe class imbalance, visual similarity between findings, and scarce expert-annotated data. This study benchmarked convolutional, Transformer, hybrid, and foundation models for multi-label bronchoscopic finding classification. Methods: ResNet-50, EfficientNet-B3, ViT-B/16, Swin-T, and Hybrid-EffSwin (average-pooling fusion of EfficientNet-B3 and Swin-T) were trained on 23 finding categories of the Bronchoscopy Examination Report Dataset (BERD) using identical hyperparameters and patient-level data splits. ConvNeXt-T and two foundation models (DINOv2 ViT-B/14 and BiomedCLIP ViT-B/16) served as additional baselines. Ablations examined fusion strategy, loss function, and decision thresholds. Models were compared using patient-level bootstrap confidence intervals (CIs) and Holm-corrected Wilcoxon tests on patient-level summaries. Results: On the independent test set (316 images, 180 patients), Hybrid-EffSwin achieved the highest macro F1-score (0.376), Matthews correlation coefficient (0.339), and area under the ROC curve (0.803). Across three training seeds, Hybrid-EffSwin ranked first in every run (macro F1 = 0.363 \((\pm)\) 0.013); its advantage over ResNet-50 and ViT-B/16 was significant in all runs, over EfficientNet-B3 in one of three, and over Swin-T in none (\((\Delta)\)F1 = +0.020 to +0.025).Among additional baselines, DINOv2 performed comparably to Hybrid-EffSwin (Hybrid-EffSwin minus DINOv2: \((\Delta)\)F1 = +0.013; 95% CI \((-)\)0.025 to 0.052) and achieved the best exact-match accuracy and calibration, whereas ConvNeXt-T was significantly inferior; a late-fusion ensemble of separately trained EfficientNet-B3 and Swin-T performed close to Hybrid-EffSwin. Average-pooling fusion performed at least as well as residual-projection and concatenation fusion, and validation-selected thresholds changed macro F1 only modestly. Conclusions: Hybrid CNN-Transformer, hierarchical Transformer, and self-supervised foundation models formed the strongest tier for multi-label bronchoscopic classification and could not be distinguished on the current test set. Multi-institutional validation is warranted.

Research Square
Gaziantep University (TR)
AI in cancer detection
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon