科研之友
微信
新浪微博
Facebook
分享链接
摘要
The conventional upper gastrointestinal endoscopy, the most commonly used method of diagnosing early gastrointestinal cancers, mainly depends on the experience of digestive endoscopists to detect the suspicious lesion. This study developed a computer-aided automatic method to detect lesion regions of early gastrointestinal cancer in gastroscopic images. First, we developed a technique to eliminate the useless regions from a gastroscopic image for enhancing the accuracy of suspicious lesion detection. Then Canny-dilation algorithm was applied to segment the image into smaller regions according to different image features. Next, we combined multiple kinds of features (including color histogram, color moment and rotation invariant local binary patterns) to describe each partitioned region. Therefore the optimal features can be selected according to the respective classification accuracy of each feature between lesions and normal regions. Finally, we detected the early gastrointestinal cancer lesions based on the optimal features in a region and used the clinical data of 407 gastroscopic images to testify the validation of the proposed method. The experimental results indicated the proposed method has the sensitivity, specificity, accuracy and G-mean of 0.93, 0.83, 0.85 and 0.88 in stomach and those of 0.83, 0.88, 0.87 and 0.86 in esophagus, and has low computation complexity. The performance of the proposed method is better than co-occurrence matrix algorithm and color wavelet covariance method treating the same data set. This pilot study shows that our established method has high sensitivity and specificity in the detection of the early gastric cancers and can provide worthy guidance to the further clinical trials.
