Related Experiment Video
Updated: Jun 24, 2025
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
DistilIQA: Distilling Vision Transformers for no-reference perceptual CT image quality assessment.
Maria Baldeon-Calisto1, Francisco Rivera-Velastegui2, Susana K Lai-Yuen3
1Departamento de Ingeniería Industrial and Instituto de Innovación en Productividad y Logística CATENA-USFQ, Universidad San Francisco de Quito USFQ, Quito, 170157, Ecuador; Colegio de Ciencias e Ingenierías "El Politécnico", Universidad San Francisco de Quito USFQ, Quito, 170157, Ecuador.
DistilIQA, a novel network for no-reference computed tomography (CT) image quality assessment, accurately predicts scan quality without a reference image. This method enhances healthcare efficiency by optimizing radiation dose and improving image analysis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- No-reference image quality assessment (IQA) is crucial for medical imaging, enabling objective evaluation without pristine images.
- Automated IQA for CT scans aids in optimizing radiation dose and healthcare efficiency.
Purpose of the Study:
- To introduce DistilIQA, a distilled Vision Transformer network for no-reference CT image quality assessment.
- To enhance network performance and efficiency through a novel two-step distillation methodology.
Main Methods:
- DistilIQA integrates convolutional operations with multi-head self-attention in a Vision Transformer architecture.
- A two-step distillation process involves a teacher ensemble network and a student network trained on predicted labels.
Main Results:
- DistilIQA achieved superior performance in quality score prediction for low-dose CT scans compared to existing CNNs and Transformers.
- Experimental analysis confirmed the effectiveness of integrating convolutional operations and the distillation approach.
Conclusions:
- DistilIQA offers a robust and efficient solution for no-reference CT image quality assessment.
- The proposed method shows significant potential for improving medical image analysis and healthcare outcomes.
