Research On Deep Learning-based Medical Image Segmentation And Its Application

With the development of medical imaging technology,doctors need to accurately and quickly extract regions of interest(ROI)from medical images when diagnosing a variety of diseases,and assist diagnosis through the changes of its shape,texture and volume.For the extraction of ROI,manual segmentation by professional doctors can ensure accurate results,but this method is time-consuming and labor-intensive,and the repeatability is poor.In order to meet the actual needs,researchers began to explore automatic segmentation methods,such as feature engineering-based and atlas-based segmentation methods,and achieved some results.Deep learning-based automatic segmentation has become a sought-after research area in medical image segmentation in recent years,driven by the advancements in deep learning technology.Taking hippocampal segmentation and leg length measurement as examples,this thesis studies the algorithm and application of medical image segmentation based on deep learning,the specific work is as follows:1.Integrating hippocampal gray matter into a cascaded deep learning framework improves hippocampus segmentation.A robust cascaded deep learning framework with integrated hippocampal gray matter(HGM)probability map was developed to improve the hippocampus segmentation(called HGM-c Net)due to its significance in various neuropsychiatric disorders such as Alzheimer’s disease(AD)and schizophrenia.Particularly,the HGM-c Net cascaded two identical convolutional neural networks(CNN),where each CNN was devised by incorporating Attention Block,Residual Block,and Drop Block into the typical encoder-decoder architecture.The two CNNs were skip-connected between encoder components at each scale.The adoption of the cascaded deep learning framework was to conveniently incorporate the HGM probability map with the feature map generated by the first CNN.Experiments on 135 T1-weighted magnetic resonance imaging scans and manual hippocampal labels from publicly available ADNI-Har P dataset demonstrated that the proposed HGM-c Net outperformed seven multi-atlas-based hippocampus segmentation methods and six deep learning methods under comparison in most evaluation metrics.The Dice(average>0.89 for both left and right hippocampus)was increased by around or more than 1% over other methods.In addition,the stability,conveniences and generalizability of the cascaded deep learning framework with integrated HGM probability map in improving hippocampus segmentation was validated by replacing the proposed CNN with 3D-UNet,Atten-UNet,Hippo Deep,Quick Net,Deep Harp,and Trans BTS models.2.Leg bone segmentation and automated measurement of leg length discrepancy from infancy to adolescence based on cascaded LLDNet and comprehensive assessment.Leg length discrepancy(LLD)is an orthopedic problem.Particularly,LLD greater than5 mm could have an increased risk of hip and knee.Therefore,accurate and reliable LLD quantification is crucial for planning appropriate treatment.Deep learning(DL)has been suggested for automated measurement of LLD on radiographs.However,these studies were only performed on children from a preschool stage and with a relatively small cohort,which is not suitable for the whole stage of children’s growth and development and various conditions,resulting in inability to meet clinical needs.The purpose of this study was to develop a unified solution using deep learning for both automated LLD quantification and comprehensive assessment on a large and comprehensive radiographic dataset covering children at full stages from infant to adolescent and with a wide range of diagnoses.Firstly,cascade LLDNet was designed to segment bilateral femur and tibia.Followed by a postprocessing of maximum connected component analysis.The leg length was calculated by localizing anatomical landmarks and computing the distance.Finally,Paired Wilcoxon signed-rank test was employed for statistical comparison,and the sensitivity and specificity of LLD greater than 5mm were also calculated.The results of experiments based on 972 children show that,the cascaded LLDNet has better performance in terms of similarity(Dice>0.97,Jaccard>0.95,Precisiosn>0.97,Recall>0.97)and stability(MD<0.23,HD<2.5,HD95<1.2,ASSD<0.25,RMSD<0.55)in leg segmentation.Regarding the LLD,high Pearson(0.94)and Concordance(0.94)Correlation values between radiology reports and automatic measurements was observed.Besides,a comprehensive assessment in terms of similarity,stability,consistency is essential in computer-aided LLD quantification of pediatrics on radiograph.