Multi-lesion classification of WCE images based on deep sparse feature selection and feature fusion

• Published in: 2022 3rd International Conference on Electronic Communication and Artificial Intelligence (IWECAI) pp. 461 - 466
• Main Authors: Lai, Zhiqiang, Jia, Zhiwei
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.01.2022
• Subjects: artificial feature; Bayes methods; capsule endoscopy; Classification algorithms; classification fusion; Cost function; depth feature; Endoscopes; Feature extraction; feature selection; Image color analysis; Robustness
• DOI: 10.1109/IWECAI55315.2022.00095

Due to a large number of capsule endoscopy (CE) images, it is extremely challenging to classify lesions directly from the images. Convolutional Neural Network (CNN) has been widely used in medical image processing. However, restricted by the limited number of labeled samples, CNN's multi-lesion detection results are lower than expected. A new method based on the fusion of Inception V3depth features and artificial features is proposed. A fine-tuned Inception V3depth model is constructed and the depth features are extracted. At the same time, artificial features such as color moments, high-order local autocorrelation (HLAC), and local binary patterns (LBP) are extracted. An improved alternative directional multiplier method (ADMM) algorithm is designed to select effective features, and single classifiers for every lesion are designed and integrated based on Bayesian Formula. The results demonstrate the effectiveness of our proposed method, 1000 images for 5 types were classified with an accuracy of 87.95% in total which is better than the state of the art.