Investigating the significance of color space 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. An experienced clinician takes 2-3 h for a complete examination of approximately 57,000-1,00,000 images received during the procedure. To reduce this time, deep neural networks/models/architectures are being explored for developing computer aided diagnosis systems for abnormality detection in wireless capsule endoscopy images, which are taken in poor illuminated environment. Though colour and texture plays an important role in highlighting the features that help in abnormality detection; existing deep neural network models consider all the images in default RGB format. Existing deep learning models for this task are highly complex and computationally extensive due to having millions of training parameters. Thus, the presented work proposes a CNN based framework, TICT-CNN (Transforming Input Color Space in Tandem with Convolutional Neural Network) for binary classification of WCE images. The proposed TICT-CNN framework initially performs data augmentation and color space conversion on-the-fly before the CNN is trained for binary classification of images. The performance of different colour spaces has been thoroughly analyzed for classification of WCE images through objective parameters and feature maps. All the experimental analysis and ablation study has been done on a real dataset obtained from All India Institute of Medical Sciences, Delhi (AIIMS Delhi). A crossdataset analysis has also been performed on a standard KID dataset. Analysis performed on real data obtained from AIIMS Delhi depicts that the performance of HSV colour space outperforms state-of-the-art approaches. The performed analysis also reflects that a simple yet efficient step of choosing the right color space can reduce the trainable parameters by 2 to 6 times and can lead to diagnosis time of only 0.02 s/frame.