Efficient SqueezeViT: A lightweight vision transformer framework for chest X-ray image classification.

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Deep Learning Architectures

Background:

• Traditional Vision Transformers (ViT) are computationally intensive for medical image analysis.
• Chest X-ray (CXR) classification requires efficient models for clinical deployment.
• Existing models may struggle with resource constraints in real-world healthcare settings.

Purpose of the Study:

• Introduce SqueezeViT, a compact Vision Transformer architecture for efficient CXR image classification.
• Address the computational demands of ViT models in medical imaging.
• Improve classification performance while reducing model size and memory consumption.

Main Methods:

• Developed SqueezeViT, a novel architecture featuring a squeezing procedure to reduce token dimensions.
• Evaluated SqueezeViT on NIH Chest X-ray and CheXpert datasets.
• Compared SqueezeViT against baseline MobileViT and other state-of-the-art (SOTA) models, including Convolutional Neural Networks (CNNs).

Main Results:

• SqueezeViT achieved significant reductions in parameters: 43.2% vs. MobileViT and up to 95.4% vs. other SOTA models.
• Demonstrated improved classification performance, with up to 16.5% increase in Area Under the Receiver Operating Characteristic Curve (AUROC) compared to SOTA.
• Outperformed baseline models and effective CNNs across various CXR classification tasks.

Conclusions:

• SqueezeViT presents a lightweight and effective architecture for CXR classification, outperforming current SOTA models.
• The model's efficiency makes it suitable for real-world clinical environments with constrained computational resources.
• SqueezeViT offers a promising solution for deploying advanced deep learning in medical diagnostics.