SALW-Net: a lightweight convolutional neural network based on self-adjusting loss function for spine MR image...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of intervertebral discs and vertebrae in spine MRI is crucial for diagnosing lumbar disc herniation.
• Convolutional Neural Networks (CNNs) are effective for segmentation but often demand high computational resources.
• Lightweight CNNs are needed for medical environments with limited computing power, but unbalanced pixel distribution in spine MRIs poses segmentation challenges.

Purpose of the Study:

• To propose a lightweight CNN, SALW-Net, for improved spine image segmentation.
• To address the sub-optimal segmentation results caused by unbalanced pixel distribution in spine MR images.
• To develop a CNN with a self-adjusting loss function to enhance learning from unbalanced pixel data.

Main Methods:

• Developed SALW-Net, a lightweight CNN incorporating a self-adjusting loss function.
• The self-adjusting loss function dynamically modifies loss weights during training to improve segmentation of unbalanced pixels.
• Evaluated SALW-Net on two distinct spine MRI datasets.

Main Results:

• SALW-Net demonstrated superior segmentation performance compared to U-net, achieving a higher average Dice Similarity Coefficient (DSC) score (0.8781 vs. 0.8482).
• SALW-Net utilizes only 2% of the parameters compared to U-net, highlighting its lightweight nature.
• Practicality validation, including deployment on lightweight devices and integration into an aid diagnosis algorithm, is underway.

Conclusions:

• SALW-Net offers a computationally efficient and effective solution for spine image segmentation.
• The proposed self-adjusting loss function successfully enhances the model's ability to handle unbalanced pixel data.
• SALW-Net shows significant clinical potential for assisted diagnosis of lumbar disc herniation, particularly in low computational power scenarios.