Related Experiment Video
Updated: Mar 3, 2026
04:25
Author Spotlight: Bridging Gaps in Anatomy and Establishing a Foundation for Algorithmic Studies
Published on: December 15, 2023
3.9K
Improving automated multiple sclerosis lesion segmentation with a cascaded 3D convolutional neural network approach
Sergi Valverde1, Mariano Cabezas1, Eloy Roura1
1Research institute of Computer Vision and Robotics, University of Girona, Spain.
Neuroimage
|April 25, 2017
Summary
We developed an automated White Matter (WM) lesion segmentation method for Multiple Sclerosis (MS) using cascaded 3D convolutional neural networks (CNNs). This approach excels with limited labeled data, achieving top rankings on public challenges and improving accuracy on clinical MS datasets.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Neuroscience
Background:
- Multiple Sclerosis (MS) is a demyelinating disease affecting the central nervous system.
- Accurate segmentation of White Matter (WM) lesions is crucial for MS diagnosis and treatment monitoring.
- Manual lesion annotation is time-consuming and subjective, necessitating automated methods.
Purpose of the Study:
- To introduce a novel automated method for White Matter (WM) lesion segmentation in Multiple Sclerosis (MS) patient Magnetic Resonance Imaging (MRI) data.
- To develop a cascaded 3D Convolutional Neural Network (CNN) architecture capable of learning from limited labeled datasets.
- To evaluate the performance of the proposed method against state-of-the-art techniques on public and private clinical datasets.
Main Methods:
- A cascaded architecture of two 3D patch-wise Convolutional Neural Networks (CNNs) was employed.
- The first CNN identifies candidate lesion voxels, while the second refines these predictions.
- The method was trained on a small set of labeled MRI data (n≤35) and evaluated on the MICCAI2008 challenge dataset and two private clinical datasets.
Main Results:
- The proposed method achieved the top rank on the MICCAI2008 challenge, outperforming 60 other methods when using T1-w, T2-w, and FLAIR MRI modalities.
- It secured 3rd position using only T1-w and FLAIR modalities.
- On clinical MS data, the method demonstrated significant accuracy improvements in WM lesion segmentation, with high correlation (r≥0.97) to lesion volume.
Conclusions:
- The cascaded CNN approach offers an effective solution for automated WM lesion segmentation in MS.
- The method demonstrates robust performance even with limited labeled training data, addressing a key challenge in medical image analysis.
- This automated technique shows promise for improving the efficiency and accuracy of MS lesion assessment in clinical practice.
