International Congress of Radiology 2023, Al-Ghardaqah, Egypt, 15 - 17 March 2023, pp.104
We aimed for automatic segmentation of parotid gland tumors using magnetic resonance
imaging (MRI) with deep learning methods.
Neck magnetic resonance imaging of patients who were operated on for parotid tumors at our
hospital between 2016 and 2022 was used. All images were obtained with a 1.5 T MRI scanner.
After excluding patients under 18 years of age and with motion artifacts, a total of 80 patients
were included in the study. Among the MRI sequences obtained, T1- weighted (T1w), T2-
weighted (T2w), and contrast-enhanced T1-weighted (CE T1w) sequences were used for
segmentation (240 images in total). The borders of the parotid gland tumors were drawn with the
joint decision of 2 radiologists, and manual segmentation was performed. Subsequently, we
prepared the data set by cropping the MR images used in the study in a standard way. Herein, 70
of each sequence image were used in training and 10 in testing. We trained three different
ResNet18-based DeepLab v3+ deep learning architectures for T1w, CE T1w, and T2w using
these MR images.
The results showed that the tumor accuracy and IoU values of ResNet18-based DeepLab v3+
deep learning architectures trained using T1w, CE T1w, and T2w test MR images were equal to
0.83176-0.80376, 0.9185-0.83463, and 0.77275-0.60153, respectively. In light of this
information, although it has been seen that the best test results have been obtained from CE T1w
images, it has been indicated that there is competition between T1w and CE T1w images.