Neural Network Powered Microscopic System for Cataract Surgery
Neural Network-Powered Microsurgical System: Advancing Precision in Cataract Surgery
Academic Context and Research Problem
Cataracts are the leading cause of blindness worldwide. Phacoemulsification combined with intraocular lens (IOL) implantation has emerged as the primary treatment method. This approach not only significantly improves patients’ visual quality but also effectively reduces the incidence of surgical complications. However, the success of the surgery heavily depends on fine operations and accurate spatial positioning and orientation of the eye. Key factors such as the location of corneal incisions, the size and position of capsulorhexis, and the angle of the IOL are critical for postoperative visual recovery.
Current ophthalmic surgical microscopes mostly rely on surgeons’ expertise and manual markings. This method faces numerous challenges, particularly in complex clinical scenarios, such as eye rotation, incomplete visual fields, corneal deformation, or external obstructions. Although commercially available microscopic navigation systems have made some advancements, they still face limitations in achieving real-time, high-precision navigation and adapting to diverse clinical conditions. Therefore, combining artificial intelligence (AI) to achieve real-time, precise navigation and assistance has become an urgent issue in the field of ophthalmic surgery.
In addressing these challenges, the authors of this paper proposed a neural network-based intelligent navigation microsurgical system designed to provide safer, standardized surgical solutions for cataract procedures.
Paper Source
The research paper, titled “Neural Network Powered Microscopic System for Cataract Surgery,” was authored by a group of researchers from multiple institutions, including Yuxuan Zhai, Chunsheng Ji, Yaqi Wang, and others, affiliated with the University of Electronic Science and Technology of China, Southwest Medical Systems, and the Chinese Academy of Medical Sciences. It was published in the February 2025 volume of Biomedical Optics Express.
Research Workflow and Innovations
Research Workflow and Technical Implementation
The core of this research is to develop an AI-based ophthalmic microscopic navigation system. The study consists of four main phases: hardware modification, algorithm design, data generation and model training, and experimental evaluations.
Hardware System Development
The research team modified a traditional surgical microscope (Zeiss OPMI Pico) by adding multiple modules, including a video recording module, an AI navigation processing module, and a projection display module. The upgraded microscope could record real-time surgical videos, process image data using deep learning algorithms, and project navigation information onto a display screen for the surgeon.Navigation Algorithm Design
The study proposed a novel end-to-end deep neural network named EyeNavNet, which achieves the following core functionalities:- Eye Center Localization: Utilizing a convolutional neural network (CNN) encoder and global feature extraction branch, it accurately completes partial visual fields during surgeries, extracts boundaries, and locates the eye center.
- Eye Rotation Tracking: By designing a rotation tracking module based on a Siamese network and correlation filters, the system could achieve multi-point non-rigid registration of eye rotation and adapt to deformations caused by surgical tools.
- Preoperative/Intraoperative Image Alignment: Using the Spatial Transformer Network (STN), intraoperative microscopic images are seamlessly aligned with preoperative slit-lamp images, ensuring precise IOL implantation parameters.
To overcome the limitation of training data for complex scenarios, the team designed a data augmentation method that simulates various surgical interferences, such as random occlusions and brightness adjustments, generating diverse data for model training.
Data Generation and Annotation
The authors collected and annotated 100 cataract surgery videos (nearly a million frames), encompassing a wide range of imaging conditions and complex surgical scenarios. For segmentation tasks, a U-Net model was fine-tuned to auto-generate initial segmentation labels, which were then manually verified for quality. For tracking tasks, an Efficient Convolution Operators (ECO) tracker was used to automatically annotate tracking points.Performance Evaluation and Simulation
- Data Scope: The study included real-world surgical data from different hospitals, divided into training (30 cases), validation (10 cases), and test sets (60 cases). The test set notably included 30 external hospital cases to assess the model’s generalization ability.
- Surgical Simulation: Eye models were used to simulate intraoperative scenarios, including eye movement and interference, to test the system’s real-time performance and accuracy.
Research Findings and Data Analysis
Improved Accuracy for Eye Localization and Rotation Tracking
- Eye center localization using the EyeNavNet algorithm achieved a positioning error (PE) of 0.121±0.044 mm, significantly better than traditional U-Net and Kernelized Correlation Filter (KCF) methods (PE: 0.160±0.129 mm).
- For eye rotation tracking, the system achieved a rotation error (RE) of 1.07±0.50°, outperforming state-of-the-art algorithms such as Ostrack, which had an RE of 1.42±0.84°.
Generalization across Datasets
During testing on an external dataset from a different hospital, EyeNavNet demonstrated robust performance, adapting well to variations in brightness, ocular deformation, and complex surgical interferences.Potential Advantages over Commercial Systems
The paper compared EyeNavNet with existing commercial systems (Callisto Eye, Verion Image Guided Systems) and highlighted its lower localization error and faster processing speed (26.2 ms per frame). Furthermore, the system’s cost-effective deployment makes it a practical option.Real-Time Performance Validation
In simulations using eye models, the system demonstrated a processing speed of 25 frames per second (fps) on a mid-tier GPU (NVIDIA GTX 1070 Ti), capable of meeting real-time surgical requirements.
Conclusions and Implications
The research team validated their AI navigation system for label-less accurate positioning and alignment of IOL during surgeries. The study not only achieved innovations in algorithm design to address limitations of existing systems but also showcased the potential for more cost-efficient system integration.
This system provides surgeons with precise and reliable navigation support in complex cataract surgeries and has the potential to expand to more advanced applications, such as multifocal IOL implants and femtosecond laser-assisted technologies. Additionally, these advancements could play a pivotal role in automating and standardizing surgical processes, further driving the digital transformation of ophthalmology.
Highlights and Future Directions
- Technological Innovations: The end-to-end EyeNavNet model significantly improves segmentation and tracking accuracy. The custom Corneal Limbus Deformation Compensation Network (LDCN) ensures stable non-rigid registration.
- Big Data Support: First large-scale surgical image dataset with diverse data augmentation strategies for complex scenarios.
- System Integration: A comprehensive hardware upgrade ensures real-time operation accelerated by GPU optimization.
Despite its promising contributions, the research has room for improvement, such as further reducing eye center localization errors in complex scenarios, validating the system on larger international datasets, and assessing its clinical impact through prospective studies.
This breakthrough signifies a leap forward in intelligent microsurgical navigation systems, with profound implications for the integration of AI in healthcare.