TFKT V2: task-focused knowledge transfer from natural images for computed tomography perceptual image quality assessment

Kazi Ramisa Rifa, Md Atik Ahamed, Jie Zhang, Abdullah-Al-Zubaer ImranJournal of Medical Imaging 2025 · DOI

TL;DR

This paper introduces TFKT V2, a novel deep learning approach for automatically assessing the perceptual quality of CT scans without needing reference images. By cleverly transferring knowledge from natural image quality assessment tasks, TFKT V2 achieves high accuracy and efficiency, closely matching human radiologist evaluations, which is crucial for balancing diagnostic reliability with reduced radiation exposure.

What Problem Does This Paper Address?

Computed tomography (CT) is a vital diagnostic tool, but it exposes patients to ionizing radiation. Minimizing this radiation is critical, yet lower doses often lead to image noise and artifacts that can obscure diagnostic details, impacting accuracy. Therefore, accurately assessing CT image quality is essential to balance radiation dose with diagnostic utility. Traditional methods for image quality assessment (IQA) are often time-consuming, require reference images (which are impractical in clinical settings), and don't always align with how radiologists perceive image quality. Existing automated IQA methods, while promising, often demand large, meticulously annotated datasets with scores that may not reflect clinical judgment. This paper aims to bridge this gap by developing a reference-free, automated CT IQA method that closely mimics radiologist evaluations, reducing reliance on extensive, clinically-aligned datasets.

Methods

The core of TFKT V2 is a task-focused knowledge transfer strategy. It employs a hybrid CNN-transformer architecture.

  1. Pre-training on Natural Images: The model first leverages a pre-trained EfficientNetV2L (a CNN) on the KADID-10k dataset. This dataset contains natural images with various artificial distortions and corresponding Mean Opinion Scores (MOS), allowing the CNN to learn general features related to perceptual image quality.
  2. Fine-tuning on CT Images: The pre-trained EfficientNet is then used as a feature extractor. This is combined with a modified Swin Transformer (Swin-KAT), which is adept at capturing both local and global image features. A "connector" module bridges these two components. This hybrid model is then fine-tuned on a low-dose CT image quality assessment (LDCTIQA) dataset.
  3. Hybrid Architecture Enhancements: The Swin-KAT backbone is further enhanced with a MLP-KAN cross-attention (MK-CA) mechanism. This combines the strengths of Multi-Layer Perceptrons (MLPs) and Kolmogorov-Arnold Networks (KANs) to improve feature representation and prediction accuracy.

This two-phase approach allows the model to learn general image quality principles from natural images and then adapt them to the specific domain of CT imaging, even with limited CT data.

Datasets

  • KADID-10k: A large dataset of natural images with 25 types of artificial distortions and human-annotated Mean Opinion Scores (MOS). Used for pre-training the CNN component.
  • LDCTIQA dataset: A benchmark dataset for low-dose CT perceptual image quality assessment, containing CT slices with varying levels of noise and artifacts, along with perceptual scores from expert radiologists. Used for fine-tuning the hybrid model.
  • Pediatric CT exam dataset: An out-of-domain clinical dataset of abdominal CT scans from pediatric patients, used for validating the model's generalizability.

Key Insights and Ideas

  • Transfer Learning from Natural to Medical Images is Viable: The paper demonstrates that knowledge learned from assessing the quality of natural images can be effectively transferred to the domain of CT image quality assessment, even though the domains are quite different. This is a key enabler for data-efficient learning in medical imaging.
  • Hybrid CNN-Transformer Architecture is Powerful: Combining the local feature extraction capabilities of CNNs (EfficientNet) with the global context understanding of Transformers (Swin Transformer) leads to a more robust and accurate IQA model.
  • MLP-KAN Cross-Attention Enhances Feature Representation: The novel MK-CA mechanism within the Swin Transformer significantly improves the model's ability to capture complex relationships and enhance feature representation for IQA.
  • Clinical Relevance is Key: The ultimate goal is to align with radiologist assessments, and TFKT V2 achieves this by directly optimizing for scores that reflect human perception, moving beyond purely technical metrics.
  • Efficiency Matters for Clinical Adoption: The model's fast inference time (around 34 ms per slice) makes it practical for real-time clinical applications.

Key Results

  • High Correlation with Radiologist Scores: TFKT V2 achieves state-of-the-art performance, with Pearson's correlation coefficients (r) of 0.9858 and 0.9861 on the LDCTIQA test set (Setting A and B, respectively), significantly outperforming many baseline and existing methods.
  • Robustness and Generalizability: The model demonstrates strong performance on in-domain data and generalizes well to out-of-domain pediatric CT exams, showing consistent diagnostic quality predictions.
  • Fast Inference: The model can assess the quality of approximately 30 CT image slices per second, making it suitable for real-time clinical use.
  • Reduced Data Dependency: The transfer learning approach allows for accurate predictions with a smaller annotated CT dataset compared to methods that require training from scratch.

Limitations and Future Work

  • Range-wise Performance Discrepancy: The model is more accurate for higher IQA score ranges (good quality images) than for lower ranges (poor quality images). This is attributed to potential differences in IQA score distributions between the training data sources.
  • Memory Consumption: While inference time is reduced, the model's memory consumption is noted as slightly higher than some top-performing LDCTIQA challenge algorithms.
  • Future Work:
    • Develop more sophisticated approaches (e.g., data shift) to address the range-wise performance differences.
  • Employ parameter-efficient model development techniques to reduce memory footprint.
  • Extend the assessment to local image quality and analyze large-scale clinical datasets across different body regions.
  • Investigate more informed sampling strategies during training for enhanced generalization.

Summary updated

Abstract

PurposeThe accurate assessment of computed tomography (CT) image quality is crucial for ensuring diagnostic reliability while minimizing radiation dose. Radiologists’ evaluations are time-consuming and labor-intensive. Existing automated approaches often require large CT datasets with predefined image quality assessment (IQA) scores, which often do not align well with clinical evaluations. We aim to develop a reference-free, automated method for CT IQA that closely reflects radiologists’ evaluations, reducing the dependency on large annotated datasets.ApproachWe propose Task-Focused Knowledge Transfer (TFKT), a deep learning–based IQA method leveraging knowledge transfer from task-similar natural image datasets. TFKT incorporates a hybrid convolutional neural network–transformer model, enabling accurate quality predictions by learning from natural image distortions with human-annotated mean opinion scores. The model is pre-trained on natural image datasets and fine-tuned on low-dose computed tomography perceptual image quality assessment data to ensure task-specific adaptability.ResultsExtensive evaluations demonstrate that the proposed TFKT method effectively predicts IQA scores aligned with radiologists’ assessments on in-domain datasets and generalizes well to out-of-domain clinical pediatric CT exams. The model achieves robust performance without requiring high-dose reference images. Our model is capable of assessing the quality of ∼30 CT image slices in a second.ConclusionsThe proposed TFKT approach provides a scalable, accurate, and reference-free solution for CT IQA. The model bridges the gap between traditional and deep learning–based IQA, offering clinically relevant and computationally efficient assessments applicable to real-world clinical settings.
BibTeX
@article{rifa2025tfkt,
  author = {Kazi Ramisa Rifa and Md Atik Ahamed and Jie Zhang and Abdullah-Al-Zubaer Imran},
  title = {{TFKT V2: task-focused knowledge transfer from natural images for computed tomography perceptual image quality assessment}},
  journal = {Journal of Medical Imaging},
  year = {2025},
  doi = {10.1117/1.jmi.12.5.051805},
}
Figure