Dilated CNN for abnormality detection in wireless capsule endoscopy images

Wireless capsule endoscopy is a non-invasive and painless procedure to examine the gastrointestinal tract of human body, and an experienced clinician takes 2-3 hours for complete examination. To reduce this diagnosis time, the present work proposes a lightweight CNN model for binary classification of WCE images. The proposed model has a strong backbone of CNN in the primary branch complemented by resolution preserving dilated convolution layers in secondary branches. The proposed model extracts multiple features at different scales and finally fuses them together to fetch the dominant global feature that aids in binary classification problem. A new dataset has been created in collaboration with All India Institute of Medical Sciences, Delhi. The efficacy of the proposed model has been verified using the developed dataset using various subjective and objective parameters. Feature maps generated at each branch have been thoroughly analyzed to understand the quality of learning. Thorough experimental analysis indicates that the proposed model yields an accuracy of 0.96, sensitivity of 0.93 and specificity of 0.97 on real data collected from AIIMS Delhi. To verify the efficacy of the proposed dilated CNN, extensive analysis has been done using standard KID dataset as well. For a fair comparison, these datasets have also been used for pre-trained inception net model. Thorough analysis indicates that the proposed architecture performs well both for AIIMS dataset and the standard KID dataset. Result analysis also reflects that the proposed dilated CNN architecture outperforms the performance of pre-trained inception net model.