Optimizing Chest X-Ray Image Classification: A Hybrid Approach Using CNN Transfer Learning and Integrated Machine Learning

優化胸部X光影像分類：使用CNN轉移學習和整合機器學習的混合方法

10.6717/JTMRT.202407_12(1).0001

Jia-Wei Lin(林家偉)；Yung-Hui Huang(黃詠暉)；Wei-Chang Du(杜維昌)；Shang-Ting Hsieh(謝尚廷)；Hsin-Yueh Su(蘇心悅)

Chest X-Ray Imaging ； Convolutional Neural Networks (CNN) ； Transfer Learning ； Machine Learning ； Diagnostic Accuracy ； 胸部X光 ； 卷積神經網絡（CNN） ； 轉移學習 ； 機器學習 ； 診斷準確性

臺灣醫事放射期刊

12卷1期（2024 / 07 / 01）

1 - 13

英文;繁體中文

Purpose: Chest X-rays are crucial in diagnosing thoracic diseases, yet their manual interpretation is prone to inconsistency and errors. This study aims to enhance the classification accuracy of chest X-ray images by harnessing the power of Convolutional Neural Networks (CNN) and integrated machine learning techniques, providing a reliable diagnostic aid for radiologists. Methods and Materials: We utilized a comprehensive dataset from the NIH Chest X- ray Database. Transfer learning was applied to fine-tune various pre-trained CNN models including EfficientNetB0, InceptionV3, MobileNetV2, ResNet101, ResNet50, ShuffleNet, and Xception. Further, features extracted from these CNNs were integrated with traditional machine learning classifiers (logistic regression, naive Bayes, SVM). The effectiveness of these approaches was compared through accuracy and Kappa index measurements, with data augmentation and sampling strategies employed to validate and generalize our results. Results: Our research indicates a significant boost in classification accuracy when combining machine learning methods, notably with the SVM classifier reaching an accuracy of 0.848 and a Kappa score of 0.837. Among CNN models, ShuffleNet and Xception showed the best performance post-transfer learning with accuracy scores of 0.716 and 0.713, respectively. Nevertheless, these CNN models were outperformed by traditional machine learning classifiers. Conclusion: The study finds that while CNN transfer learning is a promising tool for chest X-ray image classification, the integration of traditional machine learning methods with engineered features yields superior accuracy. This highlights a potential research avenue towards developing an optimal hybrid model that amalgamates the strengths of deep learning and feature engineering. The findings emphasize the need for bespoke solutions in medical image analysis and suggest a future where AI could significantly augment clinical diagnostics.

目的：胸部X光對於診斷胸腔疾病至關重要，但其人工解釋易出現不一致和錯誤。本研究旨在通過利用卷積神經網絡（CNN）和整合機器學習技術的力量，提高胸部X光影像的分類準確性，為放射科醫師提供可靠的診斷輔助。方法與材料：我們利用了來自NIH胸部X光數據庫的全面數據集。應用轉移學習對各種預訓練的CNN模型進行微調，包括EfficientNetB0、InceptionV3、MobileNetV2、ResNet101、ResNet50、ShuffleNet和Xception。此外，從這些CNN中提取的特徵與傳統機器學習分類器（Logistic Regression、Naïve Bayes、Support Vector Machine（SVM））結合。通過精度和卡帕指數測量比較這些方法的有效性，並採用數據增強和採樣策略驗證結果。結果：研究結果說明當結合機器學習方法時，分類精度顯著提升，尤其是SVM分類器達到了0.848的準確度和0.837的Kappa分數。在CNN模型中，ShuffleNet和Xception在轉移學習後表現最佳，準確度分別為0.716和0.713。然而，這些CNN模型的表現不如使用融合特徵傳的機器學習分類器。結論：研究發現，雖然CNN轉移學習是胸部X光影像分類的有效工具，但將傳統機器學習方法與CNN影像特徵結合，可以獲得更高的準確度。凸顯開發一種最佳混合模型的潛在研究方向，該模型融合了深度學習和CNN影像特徵的優勢。這些發現強調醫學影像分析中定制解決方案的需求，同時使用AI方法具有顯著提升臨床診斷可用性之價值。