Bleeding Region Segmentation in Wireless Capsule Endoscopy Images by a Deep Learning Model: Initial Learning Rate and Epoch Optimization

A common symptom in gastrointestinal tract is gastrointestinal bleeding, which can lead to serious conditions. The neural network technique is developed to segment the bleeding area in images from Wireless Capsule Endoscope. Initial variable is also importance for performance of the algorithm. In this paper, a bleeding segmentation method using a deep neural network algorithm is proposed. Variables which effect on performance of the deep learning technique in training process are studied. Initial learn rate is varied from 0.009, 0.006, 0.003, 0.06, and 0.09. Epoch is varied from 1,000 to 10,000 iterations. To evaluate the performance of segmentation method, 48 Image in KID dataset were used in the experiment. DICE rate of is 90.82%, and 69.91% for training data and test data, respectively. Based on the experiment, initial learning rate, and number of epoch effects to the performance of the method.