在醫學影像分割的任務中,神經網路可以分割複雜的醫學影像。但是它需要大量的訓練資料,且在pooling layer會丟失位置訊息,從而導致分割的精細度降低。在level set 方法中,並不需要訓練集,而且邊緣分割的結果更加精確,但是會因為雜訊的干擾而有極大的影響,並且需要手動標記位置。在本文中,我們提出了一個結合神經網路以及水平集方法(Level set method),以實現自動和準確的分割方法,用於估測超音波影片內的HA體積。所提出的方法利用從深度神經網路中學習到的影像特徵來對影像做初步的強化以及找到分割的位置,接著在用水平集方法(Level set method)來將玻尿酸切割出來。我們在超音波影片資料集上測試了我們的方法,並取得良好的結果。
In the task of medical image segmentation, neural networks can segment complex medical images. However, it requires a lot of training data, and the location information will be lost in the pooling layer, which will reduce the fineness of segmentation results. The level set method does not require training sets, and the result of edge segmentation is more accurate. But it is greatly affected by noise, and the starting position information needs to be manually marked. In this paper, we propose a combined neural network and level set method for automated and accurate segmentation. It is used to estimate the HA volume in ultrasound video. The proposed method uses the image features learned from the deep neural network to enhance the image and find the position. Then, HA is segmented using level set method and the volume is estimated. We test our method on the ultrasound video dataset and have good results.