Effects of Image Preprocessing on the Performance of Convolutional Neural Networks for Pneumonia Detection
Ercan Avşar
Abstract
Pneumonia is an infection that is typically diagnosed by taking X-Ray images of chest of the subject and it has a very high death rate in children. Recent advancements in deep learning allows for analyzing large amounts of images for detection of such diseases. In this study, effects of three image preprocessing methods on the pneumonia detection performance of deep learning models are analyzed. Contrast limited adaptive histogram equalization (CLAHE), unsharp masking, and wiener filtering are the involved preprocessing methods. For the deep learning models, two different convolutional neural network (CNN) architectures, namely, MobileNetV2 and EfficientNetB0 are trained with and without these preprocessing steps and the corresponding classification results are compared. According to the results, application of Wiener filtering operation improves the classification accuracy for both models. However, the other two preprocessing methods may improve the performance only when used with MobileNetV2. In particular, the highest accuracy, precision, and f1-score are obtained when MobileNetV2 model is trained with images preprocessed with Wiener filtering, and their respective values are 0.9199, 0.9187, and 0.9372.