Polyp is an important cause of Colorectal cancer. Wireless Capsule Endoscopy(WCE) has been widely used in direct inspection of the gastrointestinal tract without any surgical operation. Doctors need screen twenty thousand pictures per patient manually, which is time-consuming and tedious work. Recently, Deep Convolutional Neural Networks (DCNNs) shows state of the art performance in various high-level vision tasks. Therefore we present a computer-aided diagnosis model which utilises Unet to detect polyp automatically.Methods
Unet is proposed for biomedical image segmentation in recently years, showing the start of art result. In preprocessing stage, we utilise a weighted average filter to remove light spots covering polyps. Our model contains three downscale blocks and three upscale blocks. The downscale module consists of two 3*3 conv layers and a max-pooling layer which captures high-dimensional characteristics and reduces the feature maps size. In every upscale module, we add a bilinear up-sampling layer to recover the spatial information. In addition, the upscale module’s output is connected to previous downscale module’s output, which promotes integration between low-level features and high-level features and accelerates the convergence of model. In the end, we use a median filter to remove small mistake response region caused by the poor environment in the gastrointestinal tract and fill small holes using morphology.Results
We train and test our model on CVC dataset that contains pixel-level polyp segmentation label. The dataset are divided into 656 (train), 169 (validation), and 87 (test) images. The results are evaluated in terms of pixel intersection-over-union (IOU). Our method finally obtains 73.91% on IOU and operates at 23.25 fps (Frames Per Second) that is far faster than screening manually. As shown in figure 1, the result is good enough to locate polyps.Conclusions
The result shows that our method can provide efficient and accurate assistance in the diagnosis of the digestive tract, which greatly reduces the workload of doctors. It thus has a potential to apply to clinical examination.