Automating Cataract Surgery Assessment

In cataract surgery, the eye is something of a moving target. Keeping it relatively still is essential to safely and efficiently perform the procedure. Doing so requires experience and may be one way to measure surgical skill.

A newly developed artificial intelligence (AI)-powered program automatically measures surgeons’ ability to keep the eye stable, centred, and in focus during cataract surgery from video images. Using these measurements, the system also distinguishes between resident and attending surgeons with about 87% accuracy.

Resident surgeons perform as many as 150 cataract procedures during training, and many studies have shown trainees have a higher rate of intraoperative complications than experienced surgeons, noted Dena Ballouz MD. With further development, an automated system such as the one she helped develop might eventually be used to objectively assess surgical skill for training and certification.

Measuring eye movement

The AI-powered system works by analysing video images taken throughout the cataract procedure to identify three metrics for assessing eye motion: the distance between the limbus centroid and Purkinje image 1 (LCP1), a stability metric; the distance between the limbus centroid and the centre of the video frame (LCFC), a centration metric; and the focus level of the recorded video frame (FS), a focus metric. For each metric, a lower score indicates less movement and a higher level of surgical skill.

Several models were tested for each assessment component. They pre-processed images to correctly identify and measure the desired anatomical features; assessed the relationships among those features according to the proposed metrics; and assessed the skill level required to achieve a given measure.

The preprocessing component was developed with a deep-learning model trained and validated on 5,700 annotated images from 190 cataract surgeries. Then, images from 411 cataract surgeries—211 from attending surgeons and 195 from residents—were evaluated on the three proposed metrics. Finally, a skills assessment module was developed using a machine learning model to evaluate the differences in metrics between attending and resident surgeons.

Meaningful results

Using the best-performing models of each component, the case-level mean and standard deviation for all three metrics were significantly lower for attending cases than for resident cases. The three surgical steps during which residents struggled with eye stability and centration most were cortical removal, with mean LCP1 nearly 20% greater and LCP1 standard deviation more than 50% greater than for attending surgeons; viscoelastic removal, with mean LCP1 more than 50% greater; and wound closure, with LCP1 standard deviation 22% greater. Residents also struggled to maintain adequate focus throughout surgery, with mean FS scores higher than for attending surgeons across all surgical steps. In addition, the system correctly classified surgeons as resident or attending with about 87% accuracy.

“This is just the beginning of how AI can be used for cataract surgery assessment,” Dr Ballouz said. Future applications might include objective feedback—possibly in real time—for cataract surgery training and surgical skills competency tests.

Dr Ballouz presented at ARVO 2024 in Seattle, US.

Dena Ballouz MD is an ophthalmology resident at the University of Michigan, Ann Arbor, Michigan, US. dballouz@umich.edu