Differentiation Of Ocular Adnexal Lymphoma And Idiopathic Orbital Inflammation Based On Manual Features And Deep Features Of MRI

Orbital adnexal lymphoma(OAL)and idiopathic orbital inflammation(IOI)are the two most common lymphoproliferative diseases of the orbit.OAL is one of the most common malignant orbital tumors with an incidence of up to 55% and the first-line treatment for OAL is radiotherapy.IOI is an idiopathic tumor-like inflammatory disease,accounting for about10% of all orbital diseases.The mainstay of therapy for IOI is oral corticosteroids.Considering the conspicuous difference in treatment and prognosis,it is of great importance to differentiate OAL and IOI,which however remains challenging due to the similar clinical and image features.Biopsy is the gold standard to distinguish OAL and IOI,but it is an invasive diagnosis with many known complications,and depends on the accurate location of the lesion,especially the lesion in the anterior orbit and around the optic nerve,which can lead to visual impairment or even blindness if inadvertence.Therefore,it is urgent to find a noninvasive diagnostic method to distinguish OAL from IOI,which has important clinical significance.In this paper,two models based on radiomics were used to complete the differential diagnosis of OAL and IOI.The work mainly includes the following points:(1)Differentiation of OAL and IOI based on manual bag-of-features.In view of the invasiveness of the existing methods and the inability to make full use of the existing information,this study uses traditional machine learning methods to identify OAL and IOI on MRI images.Firstly,the quality of MRI images is evaluated,and the images that do not meet the inclusion criteria are discarded.Secondly,the doctors manually annotate the region of interest,and augment the training sample data.Thirdly,the region of interest is divided into patches of the same size,and each patch is extracted 160 texture features,namely manual features.Then,a bag-of-features(BOF)model is created,clustering all the training set patches to obtain BOF features with k-means++.Finally,this work trains SVM classifier and identifies the testing set data.In this study,the testing set AUC reached 0.803(95% CI:0.725-0.880).The results prove that the model based on manual features is effective and feasible in distinguishing OAL and IOI patients.Compared with biopsy and traditional methods,this method is not invasive and makes full use of existing image information,and has a high potential for clinical auxiliary diagnosis.(2)Differentiation of OAL and IOI based on deep learning and multimodal image.In view of the increasing richness of multimodal data and the possibility of providing complementary information of each mode,this study based on deep learning and fuses multimodal MRI data to identify OAL and IOI.First of all,images of all patients were evaluated,and all images corresponding to patients with missing information were discarded.Secondly,the regions of interest segmented from T1WI+C data are directly mapped to T1 WI and T2 WI images to reduce the error of manual segmentation.Thirdly,the convolutional neural network of each mode is pre-trained to get the initialization parameters.Then,initialization parameters and labels were used to extract depth features,and interexpression module was added to fuse different modal features.Spectral clustering was performed on the fused features to get the final deep features.At the end,logistic regression was used for classification.In this study,the AUC result of testing group was 0.835(95%CI :0.812-0.853).The results proved that multimodal data fusion based on deep learning can provide complementary information and consistent information,which provides a new idea for the identification of OAL and IOI.