Lmm-iqa: image quality assessment for low-dose ct imaging

Abstract

Low-Dose Computed Tomography (LDCT) reduces the amount of ionizing X-rays to which the patient is exposed; however, the accompanying increase in noise, blurring, and loss of contrast leads to a degradation in the diagnostic quality of the images. Therefore, consistency and robustness in image quality assessment have become critical requirements for clinical applications. Although numerous successful quantitative and qualitative methods are used in evaluating image quality, these methods and metrics also have various limitations. Image quality assessment approaches more commonly used in clinics rely heavily on human observation and cannot provide a standard level of assessment due to inter-observer differences. This situation highlights the need for more objective and reproducible assessment strategies. In this study, an LLM-based image quality assessment system is proposed as an alternative solution to this problem. In the proposed method, a model has been developed that generates both numerical scores and textual descriptions of clinical images, taking into account degradations such as noise, blurring, and contrast loss present in the images. The model directly explains the degradations in the images with textual and numerical feedback, making quality differences more interpretable. The main purpose of this model is to provide clinicians with a consistent decision support tool and to make the evaluation process more objective. Within the proposed method, different inference strategies such as zero-shot approach, metadata integration, and error feedback are systematically examined. The integration of contextual information allows the model to interpret visual changes from a medical perspective, while feedback mechanisms enable the system to refine its scoring accuracy. By evaluating the gains provided by these strategies individually, it has been demonstrated under what conditions the model produces output more reliably. This approach bridges the gap between purely mathematical metrics and the complex semantic requirements of radiological diagnosis. By leveraging the inherent reasoning capabilities of these models, the study moves beyond simple pattern recognition to provide a context-aware analysis of medical image integrity. The results indicate that the proposed system not only provides highly consistent scores but also offers interpretable and clear analyses that remain compatible with clinical workflows. Unlike traditional models, this methodology focuses on explainability and flexibility, providing a strategic step toward minimizing clinical workload. In this respect, the system offers a practical framework for quality assessment in Low-Dose CT imaging, which can serve as a solid foundation for the integration of artificial intelligence into future clinical practice.