Auxiliary Learning and Image Enhancement Based Pipeline to Improve CNN Performance for Pneumonia Detection

Automated disease analysis has become crucial following the need for quicker yet accurate detection to assist medical experts in their diagnosis. One of the methods that uses imaging instruments to find initial symptoms, such as Chest X-rays (CXR), is medical image analysis. Recent studies focus on finding a better model to improve their detection accuracy, leading to the development of various CNN derivatives such as ResNet. However, the performance of such a method suffers from low-quality images. Many tried to use image enhancement methods beforehand to improve the quality of the image before feeding it to the classification model. Others tried to do this automatically by incorporating denoising autoencoders and similar techniques. This paper will focus on improving the detection accuracy of pneumonia CXR classification, as the disease is diagnosable with CXR and the same approaches can be applied to similar diseases. An end-to-end approach is introduced such that the method incorporates learning the underlying features exposed by image enhancements as an auxiliary objective using an autoencoder. The proposed model is trained end-to-end with a combined loss function that enables the model to classify correctly while allowing the auxiliary task to improve the classification performance. The proposed model achieved its highest performance when using the dataset that had been enhanced using Unsharp Masking (UM) as the target for the auxiliary task, reaching 98.46% accuracy, 97.88% precision, 98.25% recall, and 98.06% F1-score.