Endoscopy, as a visual extension of surgical doctors, has been widely used since World War II. Nowadays, a new generation of endoscopic imaging system combining gradient refractive index (GRIN) waveguide and convolutional neural network (CNN) image restoration technology is a new idea for achieving small-diameter, low heat rigid endoscopes. However, this emerging endoscopic imaging system faces dual image degradation from external environmental interference and internal interference of optical imaging systems. The application of optical methods and traditional CNNs cannot distinguish degradation from different sources, resulting in high image restoration costs. The scarcity of high-quality medical imaging data is difficult to reverse in the short term, so how to achieve high-quality imaging recovery under limited data volume has become a major scientific issue in this field.

Fig. 1 Random length GRIN lens image transmission system.

The article published entitled "Ultra-robust imaging restoration of intrinsic deterioration in graded-index imaging systems enabled by classified-cascaded convolutional neural networks" in Advanced Imaging proposes a novel CC-CNN architecture for segmented correction of aberration interference encountered in GRIN waveguide based rigid endoscopes. Based on a reasonable basic assumption that image virtuality and image degradation are orthogonal to each other, image virtuality and image aberration correction networks were constructed based on ResNet50 and Unet, respectively.

Fig. 2 Image restoration by the single CNN and the CC-CNN.

This cascaded structure achieved a 9.4dB improvement in peak signal-to-noise ratio (PSNR) and a 0.09dB improvement in structural similarity (SSIM) compared to a single network trained on the same batch of data. The article provides a low-cost image correction solution for a new endoscopic architecture based on GRIN waveguide, which lays an important foundation for its development in the context of relatively scarce high-quality endoscopic imaging data.

Fig. 3 Comparison of restoration results between the single CNN and the CC-CNN for virtual imaging.

Through reasonable prior knowledge, complex problems can be decomposed into combinations of simpler subproblems, reducing the dimensionality from "multiplication" to "addition". Subsequently, multiple networks were designed for each sub problem, and the overall image correction was completed through a cascaded architecture. It significantly alleviates the imaging restoration challenges in dataset constrained scenarios, especially in areas with high imaging costs such as medical imaging and published entitled " Ultra-robust imaging restoration of intrinsic deterioration in graded-index imaging systems enabled by classified-cascaded convolutional neural networks" in Advanced Imaging.