Customized Convolution Neural Network for Multi-Class Lung Abnormality Classification From Ct Images

Abstract

Automated detection of lung abnormalities has a significant role in the computer aided diagnosis of lung diseases. Recently, medical image analysis utilizes Convolution Neural Network(CNN) to improve the outcome of clinical diagnosis. In this paper, we propose customized CNN based multi-class lung abnormality classifier from CT images. The custom CNN is trained and tested using CT images showing lung abnormalities of Carcinoma, Fibrosis, Necrosis and their performance is also compared with the results using VGG16 and VGG19. It is found that the our Custom CNN shows good results for Carcinoma, Fibrosis, Healthy, Inflammation and Necrosis with classification accuracy of 0.912 compared to VGG16 and VGG19 with accuracy of 0.7435 and 0.7216 respectively. Hence, it is proven that our custom CNN can be utilized as a second opinion to radiologist expert and improving mortality rate of these lung diseases by providing class-specific treatment for the patients.