A Novel Decision Support System for Capsule Segmentation Via High Frequency Ultrasound Images
Jingwen Zhao,Huanhuan Zhai,Xiang Liu,Jialin Song,Jian Quan Zhang,Shaolin Yang,Shuo Hong Wang
DOI: https://doi.org/10.2139/ssrn.4328050
2023-01-01
Abstract:Background: Liver cirrhosis is a common chronic liver disease. It will develop into liver cancer in the late stage and has a very high mortality rate. High-quality ultrasound (abbreviated as US) examination is helpful for its early detection and treatment. As liver cirrhosis continues to deteriorate, the physical structure of the liver capsule and liver parenchyma shown in high-frequency liver US images can also change significantly. Among them, the liver capsule is gradually ruptured, thickened, and has a stepped wave shape, while the liver parenchyma is composed of fibrous tissue. The above changes are an important basis for clinical evaluation of the degree of liver cirrhosis, but the liver parenchyma also contains noisy tissues such as blood vessels, which easily affects the extraction of ultrasonic texture features.Methods: This study develops a computational framework utilizing neural network architectures that aims at segmentation. To this end, deep learning model is needed. Deep learning greatly reduces the difficulty of object detection, even for medical images with insufficient data. However, detection usually determine the category and location of ROI(region of interest) in the image, which is easily affected by surrounding noise, especially for medical images. Conversely, individual objects such as the segmented capsule, are rarely influenced by their surroundings. Due to the existence of many suboptimal local minima in general networks, the realized neural networks only achieve suboptimal performance. This paper integrates models that have converged to local minima and focus improvements in a novel single model. Then the diagnostic value of liver capsule for liver cirrhosis is observed for the sake of diagnosis afterwards.The performance achieved by the system proposed in this paper demonstrate that it achieves good effectiveness and high precision and can be effectively applied to segmentation of bar-shaped organs, such as capsule.Results: On the high-frequency liver US dataset, the Intersection over Union(abbreviated as IOU) of UNet is 93.79$\%$, and the IOU of AA-UNet after adding attention gate and hole space convolutional pooling pyramid increases by 0.64$\%$. In addition, on the high-frequency carotid US dataset, the IOU of AA-UNet is 94.14$\%$, which is 2.00$\%$ higher than that of UNet, which verifies its robustness.Conclusions: The deep learning model provides a potential method to segment bar-shaped organs.