A fully convolutional neural network for new T2-w lesion detection in multiple sclerosis

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Longitudinal magnetic resonance imaging (MRI) is crucial for diagnosing and monitoring multiple sclerosis (MS).
• New T2-weighted (T2-w) lesions on brain MRI scans serve as a key biomarker for MS progression.
• Accurate detection of these new lesions is essential for effective patient management.

Purpose of the Study:

• To develop and evaluate a fully convolutional neural network (FCNN) for automated detection of new T2-w lesions in longitudinal brain MRI.
• To improve the accuracy and efficiency of MS lesion detection compared to existing methods.

Main Methods:

• A dataset of multichannel brain MRI scans (T1-w, T2-w, PD-w, FLAIR) from 60 MS patients (36 with new lesions) acquired one year apart was used.
• A two-part FCNN was designed: the first part learned deformation fields for nonlinear image registration, and the second part detected new T2-w lesions.
• The model was trained end-to-end using a combined loss function, with performance evaluated via leave-one-out cross-validation.

Main Results:

• The FCNN achieved a mean Dice similarity coefficient of 0.83 for lesion detection with an 83.09% true positive rate and a 9.36% false positive rate.
• Segmentation performance yielded a mean Dice similarity coefficient of 0.55.
• The proposed model demonstrated significantly superior performance (p < 0.05) and faster operation compared to state-of-the-art methods.

Conclusions:

• Combining a learning-based registration network with a segmentation network effectively detects new MS T2-w lesions.
• The proposed FCNN significantly reduces false positives and operates faster than current state-of-the-art methods.
• This approach offers a promising tool for enhanced MS diagnosis and follow-up.