Xudong Guo1, Na Zhang1, Jiefang Guo2, Huihe Zhang1, Youguo Hao3, Jingqing Hang3. 1. School of Medical Instrument and Food Engineering, University of Shanghai for Science and Technology, Shanghai, 200093, China. 2. Department of Gastroenterology, Changhai Hospital, Shanghai, 200433, China. 3. Department of Rehabilitation, Shanghai Putuo People's Hospital, Shanghai, 200060, China.
Abstract
PURPOSE: To automatically and efficiently segment the lesion area of the colonoscopy polyp image, a polyp segmentation method has been presented. METHODS: An ensemble model of pretrained convolutional neural networks was proposed, using Unet-VGG, SegNet-VGG, and PSPNet. Firstly, the Unet-VGG is obtained by the first 10 layers of VGG16 as the contraction path of the left half of the Unet. Then, the SegNet-VGG is acquired by fine-tuned transfer learning VGG16, using the first 13 layers of VGG16 as the encoder of the SegNet and combined the original decoder of the SegNet. By adjusting the input size of the Unet-VGG, SegNet-VGG, and PSPNet, the preprocessed data can be correctly fed to the three network models. The three models are used as the basic trainer to train and segment the datasets. Based on the ensemble learning algorithm, the weight voting method is used to ensemble the segmentation results corresponding to single basic trainer. RESULTS: Both IoU and DICE similarity score were used to evaluate the segmentation quality for cvc300 with 300 images, CVC-ClinicDB with 612 images, and ETIS-LaribPolypDB with 196 images. From the experimental results, the IoU and DICE obtained by the proposed method for the cvc300 datasets can reach up to 96.16% and 98.04%, respectively, the IoU and DICE for the CVC-ClinicDB datasets can reach up to 96.66% and 98.30%, respectively, whereas the IoU and DICE for the ETIS-LaribPolypDB datasets can reach up to 96.95% and 98.45%, respectively. Evaluation of the IoU and DICE in our methods shows higher accuracy than previous methods. CONCLUSIONS: The experimental results show that the proposed method improved correspondingly in IoU and DICE compared to a single basic trainer. The range of improvement is 1.98%-6.38%. The proposed ensemble learning succeeds in automatic polyp segmentation, which potentially helps to establish more polyp datasets.
PURPOSE: To automatically and efficiently segment the lesion area of the colonoscopy polyp image, a polyp segmentation method has been presented. METHODS: An ensemble model of pretrained convolutional neural networks was proposed, using Unet-VGG, SegNet-VGG, and PSPNet. Firstly, the Unet-VGG is obtained by the first 10 layers of VGG16 as the contraction path of the left half of the Unet. Then, the SegNet-VGG is acquired by fine-tuned transfer learning VGG16, using the first 13 layers of VGG16 as the encoder of the SegNet and combined the original decoder of the SegNet. By adjusting the input size of the Unet-VGG, SegNet-VGG, and PSPNet, the preprocessed data can be correctly fed to the three network models. The three models are used as the basic trainer to train and segment the datasets. Based on the ensemble learning algorithm, the weight voting method is used to ensemble the segmentation results corresponding to single basic trainer. RESULTS: Both IoU and DICE similarity score were used to evaluate the segmentation quality for cvc300 with 300 images, CVC-ClinicDB with 612 images, and ETIS-LaribPolypDB with 196 images. From the experimental results, the IoU and DICE obtained by the proposed method for the cvc300 datasets can reach up to 96.16% and 98.04%, respectively, the IoU and DICE for the CVC-ClinicDB datasets can reach up to 96.66% and 98.30%, respectively, whereas the IoU and DICE for the ETIS-LaribPolypDB datasets can reach up to 96.95% and 98.45%, respectively. Evaluation of the IoU and DICE in our methods shows higher accuracy than previous methods. CONCLUSIONS: The experimental results show that the proposed method improved correspondingly in IoU and DICE compared to a single basic trainer. The range of improvement is 1.98%-6.38%. The proposed ensemble learning succeeds in automatic polyp segmentation, which potentially helps to establish more polyp datasets.