Magnetic resonance imaging (MRI) is essential in the evaluation of various orbital diseases. However, it faces a variety of technical challenges, such as motion artifacts and susceptibility artifacts [1]. Noise, which reduces tissue contrast and obscures small anatomical structures, is a major limiting factor in visualizing the orbital structures [2]. It lowers the signal-to-noise ratio (SNR), which hinders visualization of small structures such as the optic nerve and makes subtle lesions harder to detect [3]. Although recent studies have demonstrated improved detection of optic nerve lesions using different imaging techniques and sequences [[4], [5], [6], [7], [8], [9]], this region remains challenging to analyze, primarily due to image noise [3].

Recent advances in deep learning-based image reconstruction (DLBIR) techniques show promise in reducing noise, while improving image quality and reducing acquisition times. These technologies have been successfully applied to various organs, including the brain [[10], [11], [12]]. They have improved SNR, spatial resolution, and overall diagnostic quality [10].

Among MRI vendors, several DLBIR algorithms have been developed for image denoising. These methods are typically trained to operate on high-frequency components, which are most affected by noise. By selectively reducing noise while preserving anatomical signal, they improve the SNR and enhance perceived image sharpness, particularly in sequences where higher spatial resolution or accelerated acquisitions tend to increase image noise [13,14].

While DLBIR has been evaluated in brain and general neuroimaging applications, its performance in orbital MRI has not been systematically investigated. The orbit presents unique diagnostic challenges due to its small, complex anatomy and high susceptibility to motion and noise. Improved image reconstruction in this region could enhance the detection of subtle optic nerve and orbital lesions, increase diagnostic confidence, and facilitate clinical decision-making.

The purpose of this study was to evaluate the benefits of DLBIR on the quality of orbital MR images obtained at 3 Tesla (T).