چکیده مقاله
The scarcity of open access medical datasets in the field of Multiple Sclerosis MS poses a major challenge for training deep neural networks, which typically require large amounts of data Transfer learning has been widely adopted as a remedy, where models pre trained on datasets such as ImageNet are used to initialize convolutional neural networks CNNs However, these models rely on non medical data and primarily capture low level features such as edges In this study, we leverage the similarity between brain tumor MRI scans and MS lesion data to pre train a U Net model, which is subsequently fine tuned using limited MS datasets Experimental results demonstrate that our approach significantly outperforms conventional ImageNet based pre training in MS lesion segmentation Specifically, the proposed method achieves a Dice coefficient of 65 81%, compared to 61 03% obtained with ResNet50 pre training These results highlight the advantage of using medically relevant data for pre training, enabling the network to learn domain specific texture and edge features, and ultimately improving the performance of AI based diagnostic tools for MS
کلیدواژهها
نویسندگان
Mostafa Davalloo
School of Electrical Engineering, Iran University of Science and Technology, Tehran, Iran
Ahmad Ayatollahi
School of Electrical Engineering, Iran University of Science and Technology, Tehran, Iran
شیوه ارجاع
Davalloo, Mostafa and Ayatollahi, Ahmad,1404,From Brain Tumors to Lesions: Cross-Domain U-Net Transfer Learning for MS Segmentation with Limited MRI Data,27th National Conference on Electrical, Computer and Mechanical Engineering,Shirvan
ارائهشده در
مجموعه مقالات بیست و هفتمین کنفرانس ملی مهندسی برق ،کامپیوتر و مکانیک24 شهریور 1404 · شیروان