Lightweight Compound Scaling Network for Nasopharyngeal Carcinoma Segmentation from MR Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Nasopharyngeal carcinoma (NPC) is a prevalent head-and-neck cancer.
• Accurate tumor segmentation is crucial for effective NPC treatment.
• Manual segmentation is time-consuming, and existing automatic methods demand significant computational resources, posing a challenge for smaller healthcare facilities.

Purpose of the Study:

• To develop a lightweight and accurate automatic segmentation model for nasopharyngeal carcinoma.
• To address the computational limitations faced by small and medium-sized hospitals.
• To improve the accuracy and efficiency of tumor structure identification in medical imaging.

Main Methods:

• Introduction of the LW-UNet network, a novel deep learning architecture.
• Utilization of lightweight modules to construct a Compound Scaling Encoder.
• Integration of UNet architecture benefits for a balance of efficiency and precision.

Main Results:

• The LW-UNet model achieved a high Dice coefficient of 0.813.
• The model is computationally efficient, with 3.55 M parameters and 7.51 G FLOPs.
• Achieved segmentation in under 0.1 seconds on GPU, outperforming four other state-of-the-art models.

Conclusions:

• The LW-UNet network provides an accurate and computationally efficient solution for nasopharyngeal carcinoma segmentation.
• This model is suitable for deployment in resource-constrained environments like small and medium-sized hospitals.
• LW-UNet enhances the feasibility of advanced AI-driven tumor segmentation in clinical practice.