Gland Segmentation In Colon Histology Images Using The Convolutional Neural Networks

The universalization of whole slice scanning device makes it possible to digitize pathology sections, and digitized computer-aided diagnosis of pathological image is now one of the popular areas of medical images analysis. Gland is an important organizational structure, and extraction of gland structures is an important prerequisite for subsequent computer-aided diagnosis. But due to variability of gland morphology and the different image features between malignant and benign glands, the traditional segmentation methods have significant limitations. This paper presents a more common-used colon gland pathology image segmentation method based on deep learning network, by learning the different type of features from the expert dataset.Firstly, we use the Warwick-QU dataset to establish our database for colon gland segmentation. Then we illustrate the effectiveness that using a convolution neural network as a pixels classifiers by experiments. Finally we match CNN with the sliding window and morphological processing to get the final segmentation result. But the existence of redundant computation and the low utilization of spatial information is the problem of the segmentation method mentioned above. We use the convolutional encoder-decoder architecture to solve this problem. In this architecture, we only need one time calculation to get the pixel classified output. It reduces the computational redundancy and greatly improves the speed of segmentation. But the convolutional encoder-decoder architecture makes closed glands connected. To solve this problem, we use erosion operation to preprocess the experts labeled images. Then we re-train the model for segmentation, and get better results. Our methods break the limits of prior knowledge of glandular structure, and have more versatility and scalability, solving the problem of malignant glands segmentation especially.The proposed method was verified by 165 colon histology images. The measure was defined by the similarity of the segmentation results and the experts reference centerlines in evaluation framework. In the analysis part, convolutional encoder-decoder architecture for segmentation method were compared with the CNN matched with sliding window method. The results in training set show that no matter for benign glands or malignant glands, the convolutional encoder-decoder architecture for segmentation method, combined with experts labeled images preprocessing, has higher detection segmentation accuracy, and shape similarity. The results in test set also show that the convolutional encoder-decoder architecture for segmentation method, combined with experts labeled images preprocessing, gets better results, especially for benign glands segmentation. The computation time of segmentation shows that the optimized method is more efficient. By the analysis of experimental data, an intermediate step of pixel classification results can also be used to classify colon histology images.