Abstract
Purpose :
Both cavernous hemangioma and neurilemmoma are benign tumors that occur in the orbit. It is especially meaningful to differentiate them in that they are two of the most common types of orbital tumor. In clinical work, the diagnosis of these two tumors mainly depends on Magnetic Resonance Imaging (MRI). However, in many cases, the difference in performance between them on MRI is very small- only ophthalmologists or radiologists with rich experience can distinguish them, which makes the differential diagnosis of the two very difficult. The goal of this study is to devise and develop a differential diagnosis framework, which can automatically distinguish cavernous hemangioma and neurilemmoma, to improve clinicians’ accuracy of diagnosis and enable more effective treatment decisions.
Methods :
Material: The diagnosis of cavernous hemangioma and neurilemmoma is highly dependent on observing MRI images. Therefore, our study chose MRI images as research material. For reaching a wider range of application, we selected some film version of the MRI images from over 10 different centers. We then re-cut them into over 20,000 single images after scanning. Labeling: All these images were labelled to 3 types of diseases or normal; each image is decomposed into different parts depending on the anatomical knowledge and then annotated. Training: To improve the accuracy, we added eye positioning model before classification model to reduce the scope of its identification.
Results :
This differential diagnosis framework consists of 2 stages. In the first stage, it can localize the eyes ‘position with approximately 100% accuracy. As for the second one, it can effectively classify the MRI image to normal, cavernous hemangioma or neurilemmoma, and the average accuracy in this stage reaches 90%.
Conclusions :
In conclusion, there are two considerable advantages of this new framework. Firstly, it can distinguish cavernous hemangioma and neurilemmoma, which can be exceptionally useful for the doctor to determine appropriate treatments. Secondly, this framework can assist in the clinical training of junior doctors as the rare of high-grade medical resource.
We will keep on collecting new cases to improve the accuracy as well as build a medical artificial intelligence platform based on this framework. We will also concentrate the silicon test and multihospital clinical trial to improve our study in the future.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.