Data Augmentation for Improving CNNs in Medical Image Classification

Medical image is essential for physicians to diagnose diseases. And convolutional neural networks (CNNs) have gained momentum for computer-aided diagnosis (CAD) medical image. However, there are still challenges in CNNs related to a lack of data and class-imbalanced datasets. Data augmentation is utilized to address the issues mentioned above. In this study, the Pseudo-color enhancement algorithm (OCEAN and TWILIGHT), common linear transform, CLAHE enhancement, and K-means clustering algorithm were applied to pneumonia X-ray images for data augmentation. Thus, five processed datasets were finally obtained. The results indicated that five data-augmentation methods effectively improved the performance of CNNs for the pneumonia X-ray images binary classification detection, in which the CLAHE enhancement and OCEAN enhancement methods had the accuracy and F1 Score with 97.42%, 97.19%, and 97.42%, 98.36%, respectively. DenseNet121 showed the best classification performance of four CNNs utilized in this study, as evidenced by the area under the receiver operating characteristic (ROC) curves (AUC). To verify the generalizability of the two enhancement methods, CLAHE and OCEAN enhancement methods were applied to the Magnetic Resonance Imaging (MRI) dataset of Alzheimer's disease, which had the accuracy and F1 Score with 98.15% , 98.15%, and 97.22%, 97.68%, respectively. Therefore, these simple and efficient data-augmentation methods effectively improve the performance of CNNs for diagnosing medical image, which can provide theoretical reference for physicians to diagnose the above two diseases, and advance the application of clinical medicine.