基於Transformer的耳鏡影像分類與強化因果可解釋性

// Transformer-based classification with enhanced causal explainability from otoscopic images

本研究開發了基於變壓器架構的ViT與DeiT模型，用於耳膜影像分類以提升中耳炎診斷準確度｜ViT模型達到97.78%準確率及97.30%液體滯留（effusion）F1分數，優於DeiT｜結合LRP與Attention Rollout方法提升模型解釋性，強調關鍵影像特徵｜有助於提高診斷透明度及臨床決策可靠性。

// This study developed transformer-based models (ViT and DeiT) to classify tympanic membrane conditions from otoscopic images, aiming to improve otitis media diagnosis. The ViT model achieved high accuracy (97.78%) and F1-score (97.30% for effusion), outperforming DeiT. Explainability was enhanced using a hybrid fusion of LRP and Attention Rollout methods, which effectively highlighted key image features. The approach improves diagnostic transparency and supports reliable clinical decision-making.

Published 2026年7月9日