Research On Improved Fully Convolutional Neural Network And Its Application In Medical Image Segmentation

Medical images can present human body information in multiple modalities,so they are widely used by doctors when diagnosing or treating diseases.This form of recognizing target organs from medical images only by human resources not only consumes the doctor’s time and energy,but also has subjective judgment and is prone to errors.Based on this,automated reading technology is urgently needed.In recent years,deep learning has been rapidly developed in the field of computer vision,and many scholars have devoted themselves to studying medical image processing based on neural networks.Deep learning models have strong feature learning capabilities,and perform well in the diagnosis and prediction of diseases,and the detection and segmentation of lesions.In particular,in the direction of image semantic segmentation,the fully convolutional neural networks have made a lot of achievements.A series of improved fully convolutional neural networks can automatically segment specified organs with high precision.However,deep learning models are generally not universal.Therefore,we have made different improvements to the fully convolutional neural network for different organs.The experiment was carried out on the datasets of pancreas and head & neck organs.The main contents are as follows:(1)The pancreas is a small,soft and elastic abdominal organ with high anatomical variability and low tissue contrast in computed tomography(CT),which makes the segmentation task challenging.To address this challenge,we propose a dual-input v-mesh fully convolutional network(FCN)to segment the pancreas in abdominal CT images.Specifically,dual inputs,i.e.,original CT scans and images processed by a contrast-specific graph-based visual saliency(GBVS)algorithm,are simultaneously sent to the network to improve the contrast of the pancreas and other soft tissues.To further enhance the ability to learn context information and extract distinct features,a v-mesh FCN with an attention mechanism is initially utilized.In addition,we propose a spatial transformation and fusion(SF)module to better capture the geometric information of the pancreas and facilitate feature map fusion.We compare the performance of our method with several baseline and state-of-the-art methods on the publicly available NIH dataset.The comparison results show that our proposed dual-input v-mesh FCN model outperforms previous methods in terms of the Dice similarity coefficient(DSC),positive predictive value(PPV),sensitivity(SEN),average surface distance(ASD)and Hausdorff distance(HD).Moreover,ablation studies show that our proposed modules/structures are critical for effective pancreas segmentation.(2)The distribution of head and neck organs is complicated.For the task of multi-organ segmentation,we integrate the feature pyramid framework in the classic network,U-Net,and propose a feature pyramid fully convolutional neural network to perform the segmentation work of the head & neck organs at risk.While the feature pyramid framework integrates feature maps of different levels,it can also broaden the structure of the U-Net network,alleviating the problem of gradient disappearance to a certain extent.Moreover,operations such as convolution,cascading,and addition for feature maps between different levels can reduce the semantic gap between them,thereby improving the segmentation accuracy of the model.Before the feature maps of the encoder stage are cascaded with the output feature maps of FPN,we still use the attention mechanism to highlight the region of interest.In addition,we propose a multi-scale feature fusion module to improve the nonlinear transformation capability of the network.Experiments are conducted on the public dataset PDDCA and compared with other methods.We also conducted experiments using the standard U-Net network and compared the results between the two models.A series of data shows that our improved network has better segmentation capabilities.In addition,for the improved part of the network,we conducted ablative experiments to prove the effectiveness of the proposed module.