| Objective:Lesions in the Sellar region refer to those located in or centered on the Sellar region,including pituitary adenomas,craniopharyngiomas,tuberculum sellae meningiomas,Rathke’s cysts and other rare lesions.The Sellar region is in the center of the skull base,with a deep position and a narrow space,and is surrounded by important neurovascular structures.These diseases located in this area may have similar clinical manifestations,and the diagnosis of lesions mainly depends on clinical manifestations,imaging examination and endocrinology.It is relatively easy for typical lesions,but it is very difficult to diagnose diseases without typical or special manifestations.In imaging examination,brain magnetic resonance imaging(MRI)is the first choice,but there are some shortcomings,such as(1)Due to different doctors’ knowledge and experience,different doctors may have different diagnosis of the same disease,especially the lesions with similar imaging features,even give the wrong diagnosis.(2)It consumes a lot of manpower,(3)It takes longer time.Recently,machine learning,especially deep learning,has been widely used in medical image imaging.At present,most of the segmentation tasks are manual segmentation,which is time-consuming,laborious,and inefficient.Therefore,the purpose of this study is to establish the new neural network model based on MRI images,which can efficiently and automatically identify Sellar lesions,especially pituitary microadenomas.And the efficient machine learning models would be established for distinguishing common Sellar lesions and accurate diagnosis of pituitary adenoma subtypes combined with radiomics which could provide basis and reference for clinical diagnosis and treatment of Sellar lesions.Materials and Methods:Firstly,we obtained the edge contour image of Sellar lesions by Sobel operator and edge contour information of Sellar lesions by Edge UNet network.And we established three neural network models based on U-Net network,including IE3 SNet,SRF2SNet and IERF4 SNet.The improved neural network was trained,verified,and tested and compared with the existing U-Net,Deep Labv3+ and Dense Aspp models.At the same time,the accuracy of artificial recognition was compared to evaluate the performance of the model and explore the best MRI sequence to identify lesions.Secondly,the IERF4 SNet model was used to train,verify,and test the pituitary microadenoma data set,to evaluate the recognition effect of IERF4 SNet model on pituitary microadenoma,and compared with the accuracy of artificial recognition.Thirdly,we segmented the lesions and used Simple ITK software to read and standardize the images.Also,Py Radiomics 1.2.0 was used to extract Radiomics,and data enhancement,dimensionality reduction and other methods were applied for feature processing.Xgboost,SVM and LR models were used to identify four common Sellar lesions,and SVM,KNN and NBs models were used to predict the pathological classification of PAs.Five-and ten-fold cross-validation were used to evaluate the performance of the models,and the machine learning models were compared with the accuracy of clinicians in the differential diagnosis of Sellar lesions.Results:1.Three algorithm models for Sellar region lesion recognition have been successfully established,including image edge supervises semantic segmentation network(IE3SNet),same receptive field semantic segmentation network(SRF2SNet),and image edge supervised same receptive field semantic segmentation network(IERF4SNet).IE3 SNet is a semantic segmentation algorithm for Sellar lesions based on image edge supervision.In the process of up-sampling,the algorithm combined the bottom fine-grained surface information and the image edge contour information from the Edge UNet network to make up for the edge information lost in the process of feature extraction and learn the edge features of the lesions.SRF2 SNet is based on the semantic segmentation algorithm of Sellar lesions with the same receptive field.When the sampling times of the underlying information were the same,the receptive field of the underlying information was almost the same,and then the processed underlying information was fused with the high-level semantic information.IERF4 SNet integrated the advantages of IE3 SNet and SRF2 SNet,used image edge information to supervise the network,and dealt with the underlying information in the cross-layer connection.The fusion of the algorithm can help the network to achieve ideal results.Compared with the three advanced algorithms U-Net,Deep Labv3+ and Dense Aspp in the field of semantic segmentation,this paper pays more attention to the learning of lesion edge features.And the new neural networks did not lose too much detail and bring much noise in the process of down-sampling and up-sampling.Also,they improved the receptive field of the underlying information,to better realize the recognition of Sellar lesions,especially the recognition of pituitary microadenomas.2.After the lesions are identified from the image,it was necessary to identify their general categories.In this part,three common machine learning models have been successfully established to identify four common Sellar lesions.Among the models for differential diagnosis of Sellar lesions,Xgboost model performed best.The specificity and sensitivity of T1contrast-enhanced weighted sequence were higher than those of T1-weighted and T2-weighted sequences.By comparing the accuracy of Xgboost model with that of clinicians in distinguishing Sellar lesions,the accuracy reached the clinical expert’s level and was higher than that of attending physicians and residents.3.In the automatic recognition of pituitary microadenomas,we used the previously established IERF4 SNet algorithm model for testing and verification,and the results showed that the performance of T1 contrast-enhanced weighted sequence was better than that of T1-weighted and T2-weighted sequences.The mIoU value(P <0.001)and DSC value(P < 0.015)were significantly different in different MRI sequences.The accuracy of the improved depth neural network in automatically identifying pituitary microadenomas reached the clinical expert’s level,which was significantly higher than that of attending physicians and residents.4.The SVM model performed best among many models for predicting the pathological classification of pituitary adenomas.The specificity and sensitivity of T2-weighted sequence were higher than those of T1-weighted and T1 contrast-enhanced weighted sequences.Conclusion:1.The automatic recognition algorithm models of Sellar lesions have been successfully constructed,including IE3 SNet,SRF2SNet and IERF4 SNet algorithm models.IERF4 SNet algorithm model is better than IE3 SNet and SRF2 SNet algorithm model,and it has the best recognition effect in T1 contrast-enhanced weighted sequence,and its recognition accuracy reaches the level of clinical experts,which is higher than that of attending physicians and residents,especially for pituitary microadenomas.2.Xgboost model,SVM model and LR model can be used to distinguish tuberculum sellae meningioma,craniopharyngioma,Rathke’s cyst and pituitary adenomas.Xgboost model has the best differential diagnosis effect in T1 contrastenhanced weighted sequence,and its accuracy for differential diagnosis has reached the level of clinical experts,which is higher than that of attending physicians and residents.3.SVM model,KNN model and NBs model can better predict the pathological subtypes of pituitary adenomas,and SVM model has the best prediction effect on the pathological classification of pituitary adenomas in T2-weighted sequence.The improved neural network and machine learning model can not only provide valuable guidance for neurosurgeons in clinical decision-making before operation,but also reduce the workload of doctors and provide useful suggestions for patients and their families.And it can lay the foundation for the further development of the software system for the identification and assistant diagnosis of Sellar lesions. |
PDF Full Text Request