In recent years, with the rapid advancement of microsurgical techniques, ophthalmic surgical microscopes have played a crucial role in enhancing the quality of ophthalmic surgeries. However, prolonged exposure to the light source of the operating microscope during surgery can cause phototoxic damage, such as retinopathy and mild pigment disturbance in patients. Studies have shown that reducing lighting intensity can significantly lower the risk of phototoxic damage to eye tissues and reduce photophobia in patients. Therefore, performing ophthalmic surgeries under low illumination holds significant application value. With the evolution of ophthalmic surgical microscopes, ophthalmologists are no longer confined to observing surgeries through traditional microscope eyepieces. Instead, they can observe the surgical site and perform procedures via a camera and display screen, providing an opportunity to reduce lighting intensity. Under low illumination, normal images can be obtained by extending the exposure time or increasing the camera gain. However, increasing exposure time reduces the frame rate of image acquisition, while raising camera gain introduces significant electronic noise. Hence, there is a pressing need to develop an algorithm to enhance low-illumination images captured by the microscope, improving the quality of surgical images while ensuring real-time performance.

To improve the quality of low-illuminance ophthalmic images while ensuring real-time performance, we propose a low-illuminance image enhancement network based on the ophthalmic surgical microscope. The illuminance adjustment module is designed using DCE-Net as its foundation, with improvements to the high-order adjustment curve to reduce the number of parameters and operation time. The final high-order adjustment curve is directly learned by the illuminance adjustment module, replacing the iterative process in DCE-Net. A denoising module with residual connections is introduced for image pre-processing, preserving texture information while removing noise. In addition, a hue recovery module and corresponding loss function are designed to extract the hue recovery matrix and global correction values to restore the image’s hue after illumination adjustment. A series of pig-eye images are collected under low illumination conditions using the ophthalmic operating microscope. Experimental results confirm that the proposed algorithm enhances low-light ophthalmic surgical images while meeting the real-time requirements of surgical operations and mitigating issues of noise and color distortion.

First, we compare the proposed algorithm with seven other algorithms using a low-light synthetic image dataset and a low-light pig-eye image dataset. Our algorithm preserves the original hue of the image while ensuring real-time performance and noise reduction. It achieves superior enhancement results, with the enhanced images closely resembling the ground truth (Figs. 5 and 6). Compared to other algorithms, our method performs well in peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), achieving the best results. The proposed algorithm processes a 4K image in about 7 ms, meeting the real-time requirements for surgical scenarios (Table 1 and Table 2). To further validate the algorithm’s effectiveness under various illumination conditions, we enhance low-light pig-eye images captured under different lighting conditions. Our method can reduce the illumination intensity required by the ophthalmic operating microscope by about 80% (Fig. 7). Ablation experiments are conducted for each module and loss function. Removing the RD-Module and CR-Module results in decreased SSIM and PSNR, poorer image quality, and a noticeable decline in the overall visual effect (Fig. 8 and Table 3). Without any loss function constraints, the enhanced image quality deteriorates, with reductions in both SSIM and PSNR (Fig. 9 and Table 4).

In this paper, we present a low-light ophthalmic surgery image dataset and propose a low-light image enhancement network designed for an ophthalmic surgical microscope, accompanied by relevant loss functions to guide network training. The algorithm divides the low-light image enhancement task into three sub-tasks, each with clear objectives. In this network, the RD-Module removes noise from the input image, yielding a denoising low-light image. The IA-Module then adjusts the image’s illumination pixel by pixel, producing an illumination-adjusted image. Finally, the CR-Module restores the global hue of the illumination-adjusted image to generate the final enhanced image. Experimental results show that, compared to conventional ophthalmic surgeries, our algorithm reduces the illumination intensity required by the ophthalmic surgical microscope by about 80%. In both the low-light synthetic image and pig-eye image datasets, our algorithm outperforms existing methods in terms of image quality and visual effect while meeting the real-time demands of surgical operations. The algorithm processes a 4K surgical image in just 7 ms, demonstrating its advantages and practicality. However, the algorithm’s performance diminishes when dealing with uneven illumination, and its ability to recover image detail and texture remains limited. Future work will focus on further optimizing the network’s model to enhance its robustness. In addition, adaptive enhancement of low-light surgical images will be explored to meet the specific illumination needs of surgeons in different surgical scenarios.