Identification of active lesions in the brains of multiple sclerosis patients on non-contrast magnetic resonance imagingusing artificial intelligence models

Identification of Active Lesions in the Brains of Multiple Sclerosis Patients on Non-contrast Magnetic Resonance Imaging Using Artificial Intelligence Models Purpose: To evaluate the success of artificial intelligence machine learning in identifying active MS plaques in brain on pre-contrast FLAIR MRI sequence. Material and Methods: Images of 422 patients diagnosed with MS who underwent pre- and post-contrast brain MRI between January 2017 and December 2023 were retrospectively evaluated. Patients were divided into two groups based on the presence of active plaques (789 in number) and inactive plaques (693 in number) on MRI examinations. The images were resized and converted into arrays. Randomly, 80% of the created dataset was used for training and 20% for testing. Transfer learning models including ResNet50, InceptionV3, Xception, MobileNet, MobileNetV2, VGG16, VGG19, DenseNet121, DenseNet169, DenseNet201, and NASNetMobile were used iteratively with optimization algorithms such as SGD, Adam, RMSprop, and Nadam, along with activation functions relu and swish, resulting in 88 different models. Results: The artificial intelligence model created using MobileNet with the Adam optimization algorithm and relu activation function yielded the most successful result with a 74.4% accuracy rate in classifying active and inactive plaques. Conclusion: Artificial intelligence machine learning was able to predict active plaques with moderate to high accuracy using a conventional MRI sequence, the contrast-free FLAIR sequence. Although this result may not be sufficient for clinical applications, it is promising for a diagnosis protocol avoiding the use of contrast agents. Keywords: Artificial intelligence, machine learning, multiple sclerosis.