Image pre-processing: enhance the performance of medical image classification using various data augmentation technique

The demand for techniques based on computer vision are constantly increasing due to the development of techniques for decision making pertaining to medical, social and other primary disciples of day to day life. Image processing is a subset of computer vision in which the computer vision systems make use of the image processing algorithms to carry out vision emulation for recognizing objects. This study deal with the construction of convolution neural networks (CNNs) based on deep learning. It is used for classifying chest X-ray images into two classes (Normal, Abnormal) and executed on a graphics processing unit (GPU) based high performance computing platform. Medical image classification is one of the important tasks in many medical imaging applications. Tuberculosis is a communicable disease for which early diagnosis critical for disease control. Manual screening for tuberculosis identification involves a labour-intensive task with poor sensitivity and specificity. To improve diagnosis in medical images there is in need of better classification techniques. This paper proposes CNN to classify lung X-ray images with better classification accuracy and low error rate. The data available for medical image classification problems are insufficient to train accurate and robust classifier. The data augmentation technique helps to generate more new samples from the available images using label-preserving transformations. In this paper various augmentation techniques are implemented such as horizontal flips, vertical flip, rotation (fewer angle), crops, scale right and left, are used for capturing important characteristics of medical images, and they are applied to classification function. Later little work has been done to determine which augmented strategy is best for medical image classification. Here various augmentation results are compared and evaluated to show the better augmentation techniques. It is concluded that shear lead to validation accuracies of 93% and horizontal and vertical flips gives the least accuracy of 53% of accuracy.