Swin-Kat: Advancing Swin Transformer with Kolmogorov-Arnold Network for CT Image Quality Assessment

Accurate and reliable image quality assessment (IQA) plays a pivotal role in optimizing clinical diagnosis. Most of the existing deep learning models depend on proxy IQA scores of radiologists' assessments and rely on complex architectures demanding significant computational resources. However, proxy scores may not always align well with the diagnostic quality followed by clinicians, and the complex framework limits real-time application and scalability on standard clinical hardware. In this paper, we propose a novel reference-free, automated and reliable computed tomography (CT) IQA model employing a Kolmogorov-Arnold Network-based transformer framework with an attention mechanism, dubbed Swin-KAT. Extensive evaluations demonstrate the effectiveness of the proposed Swin-Kat not only in accurately predicting in-domain radiologists' assessment but also in evaluating out-of-domain clinical images of pediatric CT exams. Furthermore, Swin-KAT is capable of quantifying the quality of approximately 50 CT images per second with minimal memory consumption, outperforming existing CT IQA methods. Our code is available at this link.