Investigating efficient CNN architecture for multiple sclerosis lesion segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Automatic segmentation of multiple sclerosis lesions in MRI aids radiologists by increasing efficiency and reproducibility.
• Current segmentation methods often employ convolutional neural networks with varying architectures, yet design choices are seldom explained.

Purpose of the Study:

• To demonstrate the effectiveness of a U-net-like architecture for segmenting multiple sclerosis lesions.
• To develop an efficient and simplified neural network model for this task.

Main Methods:

• An experimental approach was used, involving modifications (mutations and deletions) to a basic U-net architecture.
• The study analyzed the impact of architectural choices on segmentation performance.

Main Results:

• The U-net architecture's efficacy stems from its encoder-decoder structure combined with long skip connections.
• A novel, minimally parameterized U-net (MPU-net) was developed, featuring fewer parameters than existing models.
• MPU-net achieved human-level segmentation of multiple sclerosis lesions using only fluid-attenuated inversion recovery images.

Conclusions:

• Empirical analysis of the U-net architecture enhanced understanding and led to the creation of the MPU-net.
• The MPU-net, significantly less complex than other models, offers a more explainable and potentially clinically adoptable solution.
• Accurate segmentation of multiple sclerosis lesions does not require overly complex deep learning models.