MSM-ViT: A multi-scale MobileViT for pulmonary nodule classification using CT images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate classification of pulmonary nodules is crucial for early lung cancer diagnosis and treatment.
• Vision Transformer (ViT) models excel at global feature extraction in natural images but struggle with small, low-resolution medical datasets.
• Traditional Convolutional Neural Network (CNN) models have limitations in capturing global context compared to ViT.

Purpose of the Study:

• To propose and evaluate a novel Vision Transformer (ViT)-based model, MSM-ViT, for enhanced pulmonary nodule classification.
• To address the challenges of small datasets and low image resolution in applying ViT models to medical imaging.
• To achieve high performance in distinguishing benign from malignant pulmonary nodules.

Main Methods:

• A hybrid CNN-ViT approach was employed to leverage the strengths of both architectures.
• Sub-pixel fusion was utilized to enhance the extraction of subtle, small-scale features.
• Dilated and ordinary convolutions were combined for multi-scale local feature extraction.
• The MobileViT module was integrated for efficient global feature extraction and spatial-level prediction.

Main Results:

• The MSM-ViT model was validated on a dataset of 442 benign and 406 malignant pulmonary nodules from the LIDC-IDRI dataset.
• The model achieved a peak accuracy of 94.04% and an Area Under the Curve (AUC) of 0.9636 after 10-fold cross-validation.
• Performance was robust across multiple validation runs.

Conclusions:

• The MSM-ViT model effectively extracts both multi-scale local and global features from pulmonary nodule CT images.
• The proposed model demonstrates comparable performance to advanced 3D deep learning models.
• MSM-ViT offers significant advantages in computational efficiency, requiring less than 1/10 of the video memory compared to conventional 3D models.