Exploring uncertainty measures in deep networks for Multiple sclerosis lesion detection and segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Deep learning excels at segmenting large pathologies in medical images.
• Accurate segmentation of small Multiple Sclerosis (MS) lesions is crucial for patient management.
• Current deep learning models lack reliability due to deterministic predictions and difficulty with small lesion segmentation.

Purpose of the Study:

• To explore uncertainty estimation for improved deep learning-based MS lesion segmentation.
• To develop and evaluate novel uncertainty measures for small lesion detection in MRI.
• To enhance clinical adoption of deep learning by providing clinicians with reliable uncertainty estimates.

Main Methods:

• Developed a 3D Convolutional Neural Network (CNN) for MS lesion segmentation.
• Integrated Monte Carlo (MC) dropout to generate four voxel-based uncertainty measures.
• Trained the network on a large-scale, multi-site, multi-scanner clinical MS dataset.
• Computed lesion-wise uncertainties by aggregating voxel-wise uncertainties.

Main Results:

• Uncertainty filtering improved True Positive Rate (TPR) and False Discovery Rate (FDR) for both voxel and lesion-level segmentation.
• Small lesions and lesion boundaries were identified as the most uncertain regions.
• Results align with known human-rater variability in MS lesion identification.

Conclusions:

• MC dropout-based uncertainty estimation enhances the performance and reliability of deep learning models for MS lesion segmentation.
• Uncertainty quantification provides valuable insights into model confidence, particularly for challenging small lesions.
• This approach holds promise for improving clinical decision-making in Multiple Sclerosis management.