Deep learning segmentation of gadolinium-enhancing lesions in multiple sclerosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Multiple Sclerosis (MS) is a chronic neurological disease.
• Gadolinium-enhancing lesions are key indicators of active MS.
• Accurate segmentation of these lesions is crucial for monitoring disease progression and treatment response.

Purpose of the Study:

• To evaluate the performance of deep learning, specifically convolutional neural networks (CNNs), for segmenting gadolinium-enhancing lesions in a large cohort of MS patients.
• To compare the segmentation accuracy using different combinations of multispectral magnetic resonance imaging (MRI) sequences.

Main Methods:

• A 3D CNN model was developed and trained on multispectral MRI data from 1006 relapsing-remitting MS patients.
• The model's input was varied across three combinations: U5 (FLAIR, T2-weighted, proton density-weighted, pre- and post-contrast T1-weighted), U2 (pre- and post-contrast T1-weighted), and U1 (post-contrast T1-weighted only).
• Segmentation performance was quantified using Dice Similarity Coefficient (DSC), True Positive Rate (TPR), and False Positive Rate (FPR), analyzed by lesion volume.

Main Results:

• The U5 model achieved an average DSC of 0.77, TPR of 0.90, and FPR of 0.23. For larger lesions (>500 mm³), performance improved to DSC 0.81, TPR 0.97, and FPR 0.04.
• The U2 model showed slightly lower average performance (DSC 0.72, TPR 0.86, FPR 0.31), with comparable results for U1.
• Segmentation accuracy decreased for smaller enhancing lesions across all input combinations.

Conclusions:

• Deep learning CNNs demonstrate excellent performance in segmenting gadolinium-enhancing lesions in MS patients, particularly for volumes ⩾70 mm³.
• Utilizing all five multispectral MRI sequences (U5 input) yielded the best segmentation results.
• These findings support the potential of AI-driven image analysis for improved MS lesion detection and management.