Segmentation of orbital and periorbital lesions detected in orbital magnetic resonance imaging by deep learning method

Introduction
Magnetic resonance imaging (MRI) has a special place in the evaluation of orbital and periorbital lesions. Segmentation is one of the deep learning methods. In this study, we aimed to perform segmentation in orbital and periorbital lesions.

Material and methods
Contrast-enhanced orbital MRIs performed between 2010 and 2019 were retrospectively screened, and 302 cross-sections of contrast-enhanced, fat-suppressed, T1-weighted, axial MRI images of 95 patients obtained using 3 T and 1.5 T devices were included in the study. The dataset was divided into 3: training, test, and validation. The number of training and validation data was increased 4 times by applying data augmentation (horizontal, vertical, and both). Pytorch UNet was used for training, with 100 epochs. The intersection over union (IOU) statistic (the Jaccard index) was selected as 50%, and the results were calculated.

Results
The 77th epoch model provided the best results: true positives, 23; false positives, 4; and false negatives, 8. The pre­cision, sensitivity, and F1 score were determined as 0.85, 0.74, and 0.79, respectively.

Conclusions
Our study proved to be successful in segmentation by deep learning method. It is one of the pioneering studies on this subject and will shed light on further segmentation studies to be performed in orbital MR images.