Retinal Optical Coherence Tomography Image-based Pituitary Tumor Screening

Pituitary tumor is one kind of benign tumors.But with the tumor growing up,it compresses the optic chiasm or optic nerve,and causes the change of visual acuity and visual field.More than 90%of the patients with pituitary tumor has monocular or binocular vision loss and even blindness.A considerable number of patients are first diagnosed with ophthalmology as the first symptom of vision loss.But there is a lack of screening for pituitary adenomas based on retinal images.It is difficult for the ophthalmologists to determine the possibility of a pituitary tumor at the first time and may delay the diagnosis and treatment.In this paper,a new method to screen pituitary tumors based on the retinal optical coherence tomography(OCT)images is proposed.When patients go to ophthalmology for ophthalmic examination due to visual impairment,the doctor can judge the possibility of pituitary tumor.In order to achieve this purpose,retinal OCT images of patients and normal humans’ were collected.Then,traditional classification method and 3D convolutional neural network(C3D)classification method were used to screen pituitary tumor.The results of them were also compared and analyzed.The first step in the traditional classification method is to preprocess images.Secondly,the volume,thickness,gray and textural features were extracted in volume of interest(VOI).Finally,Adaboost and random forest(RF)classifiers were trained based on these features to screen pituitary tumor.When using C3D for pituitary tumor screening,a double-layer C3D network with different kernals and structures adopted in the two sub-nets is built.The final loss function was a weighted sum of two sub-nets’ loss function.Then,updating parameters according to the loss function,and the final network can be used to screen pituitary tumor.These methods were tested on 70 retinal OCT images.The average accuracy rate(Acc),true negative rate(TNR),false positive rate(FPR),recall(R),precision(P)and F2 of the Adaboost and random forest classifier were 87.14%,93.54%,6.46%,82.35%,90%,0.83 and 87.14%,89.63%,10.37%,83.57%,87.24%,0.84 separately.The corresponding index values of C3D classification were 87.50%,75.00%,25.00%,100.00%,80.00%,0.95.The experimental results show that these screening methods are feasible and can be used to assist the doctors in the diagnosis of pituitary tumors.