DeepLesionBrain: Towards a broader deep-learning generalization for multiple sclerosis lesion segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurology

Background:

• Convolutional Neural Networks (CNNs) show promise for automated Multiple Sclerosis (MS) lesion segmentation.
• Current state-of-the-art methods struggle to generalize to clinical data from unseen datasets.
• Human experts can be outperformed by CNNs in controlled evaluations.

Purpose of the Study:

• To develop a novel method, DeepLesionBrain (DLB), robust to domain shift for MS lesion segmentation.
• To achieve high performance on unseen datasets, addressing the generalization gap.
• To provide a framework suitable for clinical practice.

Main Methods:

• DLB utilizes a large group of compact 3D CNNs for robust predictions.
• Hierarchical specialization learning (HSL) pre-trains a generic network, then fine-tunes locally specialized networks.
• A new image quality data augmentation technique reduces dependency on training data specificity.

Main Results:

• DLB demonstrated superior segmentation accuracy and consistency across multiple datasets (MSSEG'16, ISBI challenge, in-house).
• The method exhibited greater generalization performance compared to existing state-of-the-art approaches.
• DLB proved robust to domain shift, performing well on unseen clinical data.

Conclusions:

• DLB offers a robust and generalizable framework for Multiple Sclerosis lesion segmentation.
• The proposed method overcomes limitations of current techniques in handling diverse clinical data.
• DLB is well-suited for clinical application due to its enhanced generalization capabilities.