A comparison on the use of Perlin-noise and Gaussian noise based augmentation on X-ray classification of lung cancer patient

The use of deep learning in medical image classification has become an important study in the past few years. The proper use of this method, however, is still hindered by many problems, one of it being the imbalance of dataset available for training which resulted in small-set database. In this study, the effect of noise-based augmentation on the performance of deep learning based classification will be studied. The noises which were used for the augmentation method were Perlin-noise and Gaussian noise. The modality of medical image used in this study is X-ray. 174 X-ray images (87 cancer, 87 normal) were used in this study and will be classified by using transfer learning from previously trained deep learning architecture. The deep learning architecture used was vgg-19. The images were divided into two groups, 138 (69 cancer, 69 normal) images were used for training phase and 36 (18 cancer, 18 normal) were images used for testing phase. Three deep learning models were used for the classification tasks, the first one was retrained to classify the original images, the second one was retrained by using mix of original images and images with Perlin-noise, and the third one was retrained by using mix of original images and images with Gaussian noise. The results showed that the three models returned similar accuracy of 0.8 which indicate that the use of noise-based augmentation can increase the performance of deep learning in classifying medical images with small set training database.