Automated Enhanced Handheld Fundus Photography via Unpaired Learning

Retinal fundus photography (camera) serves as crucial instrumentation for the clinical measurement, screening, and diagnosis of both ocular and systemic diseases. While traditional table-top fundus photography is prevalent, handheld fundus photography (HFP) is critical in situations demanding mobility, such as emergency evaluations, remote field diagnostics, and neonatal screenings. Despite its adaptability, HFP’s utility is often compromised by the prevalence of low-quality images, leading to increased clinical uncertainty and misdiagnosis. To address the above concern, in this work, we propose a novel lightweight deep unpaired consistency enhancement (DUCE) pipeline to HFP, named HFP-DUCE, incorporating handheld fundus image quality assessment (HFIQA) and handheld fundus image quality enhancement (HFIQE) modules. This study not only establishes a standardized imaging protocol for HFP but also presents the first comprehensive dataset for HFP-DUCE. To make allowances for the inherent calculation limitations of HFP devices, we thoroughly benchmark the HFP-DUCE and propose a lightweight FDU-GAN for the HFIQE. Leveraging the proposed FDU-GAN, qualitative and quantitative experiments demonstrate that our method effectively enhances the quality of fundus images and achieves the state-of-the-art (SOTA) quantitative metrics with only 4.51 million parameters in the inference phase, which is 17 times lighter than the previous StillGAN. The HFP-DUCE pipeline enhances handheld fundus cameras, making them easier to use while delivering robust, high-quality (HQ) images, further facilitating downstream measurement, analysis, and accurate diagnostics.