Answers are grounded in this paper and cite their sources.
Polyps and gastroesophageal reflux disease (GERD) are among the most common gastrointestinal (GI) disorders, and together account for a substantial share of the global digestive-disease burden. Endoscopy remains the reference examination, but diagnostic yield depends heavily on the operator and early lesions are easy to miss during routine screening, making automated interpretation a compelling goal for deep learning. We evaluate eight compact convolutional and vision-transformer backbones on the four-class GastroEndoNet dataset [1] (Gerd, Gerd Normal, Polyp, Polyp Normal), a clinician-annotated set of 4,006 endoscopic images. Each backbone was trained with the same three-stage progressive fine-tuning schedule, so differences in performance reflect architecture rather than training procedure. Among pretrained networks, EfficientNet-B1 achieved an accuracy of 78.74% ± 0.33%, while MobileNetV3-Small attained a favourable AUC of 0.9444 ± 0.0025 at just 3.78 MB. Networks trained from scratch did not reach usable accuracy. The two strongest checkpoints then entered an eight-stage fine-tuning pipeline with a cumulative ablation adding, one at a time, higher input resolution, cosine learning-rate decay, stronger spatial and photometric augmentation, and categorical focal loss, revealing which components helped and which proved harmful at this data scale. Using test-time augmentation (TTA × 10) and 95% bootstrap confidence intervals, the fully configured EfficientNet-B1 reached 83.50% accuracy (95% CI [0.802, 0.864]) and AUC 0.9691 (95% CI [0.961, 0.977]), with low run-to-run variability (σ = 0.0068). MobileNetV3-Small was not far behind at 78.85% accuracy and AUC 0.9585, with just 3.78 MB and 0.94M parameters, roughly 6.7× smaller. Notably, test-time augmentation contributes the single largest accuracy gain (+6.6 points). Paired significance tests (McNemar, DeLong, bootstrap) confirm EfficientNet-B1 significantly outperforms MobileNetV3-Small (p < 0.05); a near-duplicate audit, calibration (ECE ≈ 0.03) and zero-shot external validation on HyperKvasir temper the clinical claims, which remain preliminary pending multi-centre validation.
No takes yet. Share an insight, caveat, or question.
Mahanta et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: