Could AI–Based Real-Time Surgical Guidance Reduce Complications During Cataract Surgery?

A recent study suggests a conceptual deep neural network–based surgical guidance platform is precise and has potential as a valuable tool during phacoemulsification procedures.

Despite advances in surgical technique and instrumentation, surgeons performing phacoemulsification procedures deal with variables that can negatively affect the procedure’s safety and patients’ visual outcomes. A study published in JAMA Ophthalmology investigated the potential of an artificial intelligence (AI)–based surgical feedback platform to assist surgeons and improve procedures.

Phacoemulsification is a cataract procedure in which a surgeon emulsifies the eye’s internal lens, the lens is aspirated from the eye, and then replaced with a balanced salt solution to maintain the anterior chamber of the eye. Variables such as rapid changes in intracameral flow and associated fluid turbulence, instrument positioning, and visualization of intraocular tissues can affect the procedure’s performance and outcomes.

Previously, computer vision algorithms and deep neural networks (DNNs), a type of AI, have been used for post-surgical analysis. They have been shown capable of segmenting the pupil and surgical instruments, analyzing features to determine a trainee’s performance, and identifying surgical phases in pre-recorded videos. Different convolutional neural networks (CNNs) have been up to 95% accurate, according to the study authors.

For this study, researchers developed a surgical guidance platform using a region-based CNN (R-CNN) — a subtype of DNN— and assessed the platform’s ability to track the pupil, identify surgical phases, and activate specific computer vision tools to potentially aid surgeons with audiovisual feedback in real-time during phacoemulsification surgery.

Ten surgical procedures performed by attending and trainee physicians at the University of Illinois Hospital and Health Sciences Center were recorded with a stereoscopic surgical microscope frame by frame, and 6 were selected at random to train the DNN in the following feedback and visualization tools:

  • Capsulorhexis guidance for improved symmetry and intended size of the rhexis
  • Feedback on decentration of the eye, erratic tool movement, and turbulent flow conditions
  • Enhanced visualization of anatomical structures via contrast equalization, including visualization of the rhexis, remaining lens fragments, and cortical fibers

Main outcomes were area under the receiver operator characteristic curve (AUROC) and area under the precision-recall curve (AUPR) for surgical phase classification, as well as Dice score for detection of the pupil boundary.

The DNN was trained with 600 frames and evaluated with 23,640 frames in a heterogeneous set of phacoemulsification cases. It was compared with 10,100 frames pulled from the publicly available Cataract-101 dataset to evaluate the generalizability of the platform.

The DNN achieved AUROC values of 0.996 for capsulorhexis, 0.972 for phacoemulsification, 0.997 for cortex removal, and 0.880 for idle phase recognition, researchers found. When applied to the external dataset, the mean performance decrease was 6.8%. The final algorithm reached a Dice score of 90.23% for pupil segmentation in the local dataset and 85.4% in the external dataset. The final platform had a processing rate of and a mean processing speed of 97 (standard deviation 34) frames per second.

Eleven cataract surgeons evaluated the surgical guidance platform post hoc, 8 (72%) of whom responded that they would be mostly or extremely likely to use the tool for complex procedures. Five (45%) respondents said they would consider it useful for non-complex procedures. All participants agreed the platform could be useful for real-time surgical guidance in phacoemulsification procedures, and 10 of 11 participants considered the pupil tracking and phase classification tools mostly or extremely accurate.

The results were also achieved with a relatively low number of frames for training the DNN.

“This proof-of-concept investigation suggests that a pipeline from a surgical microscope could be integrated with neural networks and computer vision tools to provide surgical guidance in real time,” the authors concluded, noting that the feasibility of such a system’s implementation requires more research to determine.

Reference

Nespolo RG, Yi D, Cole E, Valikodath N, Luciano C, and Leiderman YI. Evaluation of artificial intelligence-based intraoperative guidance tools for phacoemulsification cataract surgery. JAMA Ophthalmol. Published online January 13, 2022. doi:10.1001/jamaophthalmol.2021.5742