ALL-Net: Anatomical information lesion-wise loss function integrated into neural network for multiple sclerosis...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Accurate segmentation of multiple sclerosis (MS) brain lesions is crucial for diagnosis and treatment monitoring.
• Lesion variability in size, shape, location, and contrast presents a significant challenge for automated detection.

Purpose of the Study:

• To develop an automated algorithm, ALL-Net, for fast and accurate segmentation of MS brain lesions using deep convolutional neural networks.
• To integrate anatomical information and a novel lesion-wise loss function to improve segmentation performance.

Main Methods:

• Developed ALL-Net, a deep convolutional neural network incorporating distance transformation mapping for encoding lesion-specific anatomical information.
• Implemented a lesion-wise loss function to address class imbalance and enhance detection of smaller lesions by modeling them as spheres of equal size.

Main Results:

• On the ISBI-2015 dataset, ALL-Net achieved a top performance score of 93.32.
• On the Cornell MS dataset, ALL-Net demonstrated significant improvements in both voxel-wise (Dice, AUC) and lesion-wise (F1 score, AUC) metrics compared to existing tools.
• Statistical analysis confirmed significant improvements across various metrics (p < 0.0001 for most).

Conclusions:

• ALL-Net provides a robust and accurate solution for automated multiple sclerosis lesion segmentation.
• The integration of anatomical context and a specialized loss function effectively addresses challenges posed by lesion heterogeneity.
• This algorithm holds potential for improving clinical workflow and patient outcomes in multiple sclerosis management.