A Study On 3D Medical Image Segmentation Method Based On Improved V-Net

Medical image segmentation plays a crucial role in clinical diagnosis.In order to diagnose diseases accurately,it is necessary to manually mark the position of nodules on CT slices.However,this method is time-consuming and unstable.Many techniques have been proposed by researchers in recent years,for medical image segmentation,but these methods often fail to fully utilize global contextual information and are prone to feature loss.In this paper,we focus on lesion segmentation in CT images using deep learning methods.Our main contributions include the following.1.A novel image preprocessing method for pulmonary nodule datasets is presented.This includes a lung parenchyma segmentation method,a CT image preprocessing method and pulmonary nodule mask generation method.This method effectively removes irrelevant human tissues from CT images and addresses problems such as uneven distribution of image voxel spacing and low feature discrimination;this method has good universality and can be extended to other CT image datasets.2.A 3D medical image segmentation method based on NLA-VNet is proposed.To overcome the limitations of convolutional neural networks in obtaining long connection context information and fitting models,the channel and position attention modules are designed to be used in combination with VNet.Additionally,an improved loss function W-Dice is proposed for medical image segmentation tasks.Compared with the baseline network,this methos achieves a 4.9% increase in Dice coefficient value,2% increase in precision rate,6.5% increase in recall rate,and 4.3% increase in F1 score.The effectiveness of proposed method has been tested.3.Another 3D medical image segmentation method based on IAD-VNet is proposed.To address the problems of feature loss caused by down-sampling of the encoding/decoding architecture,insufficient attention to shallow features,and gradient disappearance during training,an iterative deep aggregation structure,deep supervision strategy,weight learning module,and incomplete dense connection structure are used to improve V-Net.The Dice similarity coefficient,precision,recall and F1 scores achieved by this method are 0.835,0.846,0.837 and 0.841,respectively.These results outperformed both the baseline network and existing state-of-the-art methods,demonstrating the effectiveness of our proposed approach.