Font Size: a A A

Bone Segmentation Methon Of Lower Limb Ct Images Based On Deep Learning

Posted on:2021-06-23Degree:MasterType:Thesis
Country:ChinaCandidate:Z B JiangFull Text:PDF
GTID:2504306107468824Subject:Computer technology
Abstract/Summary:PDF Full Text Request
Segmentation of bones from CT images is a very important process for many medical tasks.It is the basis for fracture localization and diagnosis of bone diseases,which can help radiologists make medical decisions,and also provide some stable references for analysis and segmentation of other parts of the human body(such as organs).According to the characteristics of CT data of lower extremities,a bone segmentation method for CT scan data of lower extremities is proposed based on Unet.First,the pretrained CNN on Image Net is used as the backbone of the encoder to solve the problem of weak feature extraction ability of the Unet model.Secondly,an attention mechanism is introduced to dynamically adjust the feature weights during the training process and fuse features from different pyramid ratios to increase the receptive field and solve the problem of large bone tissue differences.Then stack upsampling feature maps from all decoder blocks and use them as input to the final layer to capture more semantics while generating more precise positioning.Finally,the weighted sum of dice loss and binary cross-entropy loss is used as the final loss to effectively reduce the phenomenon of false segmentation and insufficient segmentation.In addition,in order to reduce the cost of manually labeling data,a semi-supervised method for bone segmentation of CT scan data of lower limbs is proposed,which uses unlabeled data for multiple rounds of self-training.At the same time,through the CNN integrated scheme,the error amplification during self-training is reduced.Finally,the performance and generalization ability of the newly proposed method were tested on 10 sets of lower limb CT scan data that were not involved in training.Experimental data shows that the segmentation effect of the newly proposed method is significantly improved compared with other CT methods of lower limbs,and the average dice coefficient reaches 0.98.
Keywords/Search Tags:Bone Segmentation, Deep Learning, Unet, Semi-supervised Learning
PDF Full Text Request
Related items