Efficient Pneumonia Detection via Transfer Learning: A Bayesian CNN Approach to Chest X-Ray Classification

摘 要

本研究聚焦於利用貝式卷積神經網路（Bayesian Convolutional Neural Networks, BCNN）在胸腔X光影像分類中的應用，重點探討肺炎檢測。研究採用遷移式學習和模型微調技術，將原先為多類別分類開發的預訓練BCNN模型，專門調整用於肺炎的二分類檢測。透過在新的胸腔X光影像資料集上進行微調，模型在測試資料上成功達到91.19% 的準確率，優於VGG16（86.54%）、InceptionV3（87.98%）和ResNet50（74.04%）。此外，BCNN 在 AUC（0.98）、精確率（precision）、召回率（recall）與 F1-score 方面均展現卓越性能，證實其在區分正常（Normal）與肺炎（Pneumonia）病例方面的穩健表現。展現在醫學影像診斷中的顯著潛力。遷移式學習為模型帶來多重優勢：顯著降低運算成本，加速模型收斂，並保留預訓練資料中的關鍵特徵。這種知識遷移使模型能有效適應醫學影像的特定任務，即使在有限的資料集下，依然能維持高準確性和良好的通用性。貝式框架的導入進一步提升了模型的可靠性，利用信心值估算的方法，有效減少過度擬合的風險，這在醫療應用中尤為重要。為實現臨床實際應用，研究團隊將分類模型部署於基於Flask的網頁應用程式，實現從胸腔X光影像的即時肺炎檢測。此舉使醫療專業人員可將模型作為診斷輔助工具，提供快速且可靠的分析，從而協助臨床決策。研究結果充分展示了深度學習技術，特別是結合貝式卷積神經網路與遷移式學習方法在醫學影像領域的潛力，為提升診斷準確性和效率提供了嶄新的技術路徑。

Abstract

This study explores the application of Bayesian Convolutional Neural Networks (BCNN) in chest X-ray image classification, specifically for pneumonia detection. Employing transfer learning and fine-tuning, we adapted a pre-trained BCNN model—originally designed for multi-class classification—to perform binary pneumonia classification. The model achieved a highest accuracy of 91.19%, outperforming VGG16 (86.54%), InceptionV3 (87.98%), and ResNet50 (74.04%). Additionally, BCNN demonstrated superior AUC (0.98), precision, recall, and F1-score, confirming its robust performance in distinguishing between NORMAL and PNEUMONIA cases. Transfer learning provided multiple advantages, including reduced computational costs and accelerated model convergence by leveraging pre-trained feature representations. The Bayesian framework further enhanced model reliability by providing uncertainty estimates and mitigating overfitting risks, which is crucial in medical applications with limited labeled data. We deployed the classification model in a Flask-based web application, enabling real-time pneumonia detection and providing healthcare professionals with an effective diagnostic support tool. This research highlights the promising role of deep learning technologies, particularly BCNN with transfer learning, in improving diagnostic accuracy, efficiency, and clinical usability. Future work could extend this approach to multi-class disease classification and explore its applicability in other medical imaging modalities.