Research On Semantic Segmentation Of Sellar Lesions Based On Deep Learning

The sella area is adjacent to the circle of Willis,optic nerve and other important neurovascular structures,which are the common sites for lesions.Craniopharyngioma(CR),tuberculum sellar meningioma(TSM),rathke’s cleft cysts(RCCs)and pituitary adenomas(PAs)are the most common lesions in the sellar region.As a noninvasive imaging modality,the Magnetic resonance imaging(MRI)has the advantages of high resolution and large amount of information,which is regarded as an essential examination method for accurate diagnosis of sellar lesions.At present,there are still many difficulties in accurately identifying the sellar lesions.Firstly,some small lesions in the sellar region,especially pituitary microadenomas,are very difficult to identify due to their potential artifacts and the possibility of mutation.Although dynamic MRI scans can improve the diagnostic ability of microadenomas,the false positive diagnosis are still frequently happened.Secondly,due to the high variability in the shape and structure of sellar lesions,some patients with sellar lesions may need multiple sequences or even different imaging examinations to get an accurate diagnosis.Finally,lesion diagnosis is highly depends on the expert knowledge or experience,which means there exist great differences in the diagnosis of physicians and category of diseases.Especially for lesions with similar imaging characteristics,wrong diagnosis given by physicians are possible in clinical diagnosing.These factors make it difficult to accurately identify the disease so that influence the making of treatment plan.In recent years,the computer-based automatic identification methods have shown promising ability to help the diagnosis of physicians,which could save the doctor’s time,provide accurate and reliable results,reduce the probability of missed diagnosis.In this research,with the MRI images,a deep learning-based semantic segmentation algorithm for sellar lesions was studied,and some improvements were made based on the state-of-the-art algorithms,which makes it possible to accurately identify sellar lesions.The main contributions of this research are summarized as follows:Since the existing data set could not support this research,with the help of the Department of Neurosurgery,West China Hospital of Sichuan University,we established a large-scale MRI image data set of sellar lesions containing large and small lesions.This data set contains data of 259 patients with sellar disease,including 81 CR,54 TSM,61 RCCs and63 PAs.In order to make up for the image information lost by the neural network during the down-sampling process,this research proposes a sellar lesions semantic segmentation algorithm based on image edge supervision(IE3SNet).The proposed algorithm firstly extracts the edge feature information of MRI images of sellar lesions,and then fuses the extracted edge information during the upsampling to assist neural network to recognize lesions.Compared with the three comparison algorithms in this thesis,the IE3 SNet algorithm has achieved the best segmentation result on the three sellar lesions data sets.Aiming at the problem that existing algorithms cannot well identify the small lesions,this research proposes a semantic segmentation algorithm for sellar lesions based on the same downsample frequency(SDF2SNet).The algorithm will process the information after each downsampling through an U-Net to improve the receptive field of this information to obtain multi-scale features.The SDF2 SNet algorithm integrates multi-scale feature information,and uses high-level semantic information to guide low-level fine-grained information to detect small targets.The segmentation results of the SDF2 SNet algorithm on the three test sets show that the algorithm can effectively identify small lesions in MRI images of sellar lesions.Finally,this research combines the advantages of the IE3 SNet algorithm and the SDF2 SNet algorithm,and proposes a sellar lesion semantic segmentation algorithm(IEDF4SNet)based on image edge supervision and the same receptive field.Experiments show that compared with three advanced comparison algorithms,the IEDF4 SNet algorithm can achieve the best segmentation performance on sellar lesions.At the same time,this research also implemented a sellar lesions automatic segmentation system(SALSystem)based on the IEDF4 SNet algorithm to reduce the workload of clinicians and provide potential guidance for clinicians’ preoperative decision-making.The test results show that the proposed SALS system can segment the sellar lesions more accurately,which could be promising in practical applications.