Improving MS Lesion Detection in MRI Images Through Feature Extraction from Images an‎d Deep Learning Models

9/16/2025 12:00:00 AM

Today, automatic detection of Multiple Sclerosis (MS) lesions in MRI images, especially in the FLAIR (Fluid-Attenuated Inversion Recovery) imaging sequence, is one of the important an‎d challenging issues in the field of medical image processing. Due to the variability in size, shape, an‎d pixel intensity of these lesions, as well as their similarity to other brain abnormalities, accurate diagnosis is difficult for physicians. From a technical an‎d software engineering perspective, this problem is considered as a pattern classification an‎d localization task in complex three-dimensional image data, which can be improved using deep learning methods. In this study, a lesion detection an‎d localization model based on advanced convolutional neural network architectures has been developed, utilizing modern structures such as YOLO (You Only Look Once) for high-precision detection of small objects. The proposed model aims to improve accuracy in identifying small an‎d scattered lesions by leveraging diverse, multi-scale feature extraction an‎d optimized geometric representation. In other words, the goal of the model is to predict a bounding box that has the highest overlap with the true lesion boundary, so that the closest possible box to the actual lesion area is identified. The images used are three-dimensional with the .nii format, consisting of FLAIR image volumes. The lesion annotations were performed by healthcare specialists based on pixel intensity an‎d precisely segmented such that the boundary of each lesion is clearly defined in separate images labeled as LESION. For eva‎luating the proposed model, quantitative metrics such as accuracy, mean average precision, an‎d Dice similarity coefficient were employed. Experimental results show that the proposed method performs better in detecting small lesions in various brain regions compared to state-of-the-art methods such as classic YOLO an‎d U-Net. Additionally, the execution time of the proposed method is acceptable compared to existing approaches, enabling its use in clinical applications an‎d computer-aided diagnosis systems.

بخش بندي , جعبه محدود كننده , ضايعه , ام اس , يادگيري عميق

Bounding Box , Segmentation , Lesion , Multiple Sclerosis (MS) , Deep Learning

Erfan Najafi

Dr. Einollah Khanjari