Identifying Intraoperative Floppy Iris Syndrome (IFIS) Using AI

Technical Summary: Identifying Intraoperative Floppy Iris Syndrome (IFIS) Using AI

Problem Statement
Intraoperative Floppy Iris Syndrome (IFIS) is a condition characterized by poor pupillary dilation and an elastic iris that resists mechanical stretching during cataract surgery. It significantly increases the risk of intraoperative complications, including iris injury, wound dehiscence, posterior capsule rupture, and vitreous loss, which can lead to long-term visual impairment. Currently, preoperative prediction of IFIS relies primarily on clinical risk factors, most notably the use of alpha-1-antagonist medications. However, a significant portion of IFIS cases occurs in patients without a history of such medication use, and clinical assessment alone may miss subtle anatomical indicators. The study aims to evaluate whether Artificial Intelligence (AI) algorithms can utilize routine preoperative anterior segment imaging to identify patients at risk for IFIS more effectively than current methods.

Methodology
The study utilized a retrospective cohort of 2,220 patients (3,239 eyes) who underwent cataract surgery at Rabin Medical Center between April 2016 and November 2022. The dataset was split into training (60%), validation (20%), and testing (20%) sets, ensuring that images from the same patient ID did not appear across different sets to prevent data leakage.

• Data Acquisition: Preoperative images were acquired using the Tomey OA-2000 optical biometer, which captures both infrared (IR) images and swept-source Optical Coherence Tomography (SS-OCT) images. IFIS diagnosis was confirmed based on the surgical report where the operating surgeon explicitly noted the presence of the syndrome.
• Image Processing: Infrared images (originally 360×330 grayscale) were resized to 224×224 pixels, centered on the pupil, and converted to RGB tensors. Data augmentation techniques, including horizontal/vertical flips and "Pizza-Slicing" (cropping images into 4 or 8 slices), were applied to the training set to address class imbalance.
• Model Architectures:
  • Infrared Classification: Four Convolutional Neural Network (CNN) architectures were evaluated: ConvNeXt-large-224, ResNet-50, EfficientNet-b7, and MobileNet. All were pre-trained on ImageNet-1K and fine-tuned for 20 epochs.
  • OCT Classification: A ResNet-50 model was trained on anterior segment OCT images.
  • Anomaly Detection: An EfficientAd model (student-teacher architecture) was employed to detect IFIS as an anomaly within the dataset.
  • Ensemble Strategy: For OCT images, majority voting was applied across multiple images of the same eye to determine the final classification.
• Human Baseline: Two experienced cataract surgeons, masked to all clinical data, manually assessed a subset of 100 infrared images to establish a human performance baseline.
• Statistical Analysis: Performance was evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC), sensitivity, specificity, and accuracy. Multivariate logistic regression was also performed to correlate clinical variables (e.g., alpha-1-antagonist use, pupil size, anterior chamber depth) with IFIS occurrence.

Key Results

• Clinical Demographics: The cohort included 243 eyes (7.5%) with confirmed IFIS. Significant differences were found between IFIS and non-IFIS groups regarding age (higher in IFIS), alpha-1-antagonist use (14.3% vs. 4.7%), anterior chamber depth (shallower in IFIS), lens thickness (wider in IFIS), and pupil size (smaller in IFIS).
• Multivariate Logistic Regression: A model based on clinical variables achieved an AUC of 70%. Key predictors included alpha-1-antagonist use (OR 1.14), gender, and pupil size.
• Human Graders: The two human surgeons achieved AUCs of 55% and 49% respectively when assessing infrared images alone, indicating limited ability to visually predict IFIS without additional data.
• AI Model Performance:
  • Infrared Classification: The ConvNeXt-large-224 model achieved the best performance among CNNs with an AUC of 65.5%. Other models (EfficientNet-b7, MobileNet, ResNet-50) performed lower, with AUCs ranging from 39.8% to 54.6%. The models struggled with class imbalance, showing low binary recall (e.g., 4.26% for ConvNeXt) despite high weighted recall.
  • OCT Classification: The ResNet-50 model on OCT images achieved an AUC of 57.8% on individual images, which improved to 61.7% when majority voting was applied.
  • Anomaly Detection: The EfficientAd model achieved an AUC of 53.3% (53.5% in results section), with perfect recall (100%) but low precision (26.8%) and accuracy (26.9%).

Key Contributions and Significance
The study demonstrates that AI algorithms can assist ophthalmologists in identifying patients at risk for IFIS using routine preoperative imaging, potentially outperforming experienced surgeons when relying solely on image analysis. Specifically:

1. Superiority to Human Visual Assessment: The best AI model (ConvNeXt-large-224) achieved an AUC of 65.5%, surpassing the human graders' performance (49–55%) in predicting IFIS from infrared images alone.
2. Integration of Imaging Modalities: The research validates the utility of both infrared and OCT imaging for this task, with the multivariate logistic regression (clinical data) achieving the highest AUC (70%), suggesting that AI image analysis complements rather than replaces clinical risk factor assessment.
3. Addressing Class Imbalance: The study highlights the difficulty of detecting IFIS due to its low prevalence (~7.5%) and the resulting class imbalance, which severely impacts model precision and recall. The authors note that while anomaly detection offered perfect sensitivity, it required further optimization to balance precision.

Limitations and Future Directions
The authors acknowledge several limitations, including the unbalanced dataset, the variability in surgeon reporting of IFIS severity, and the use of grayscale infrared images which lack color information (such as iris color, previously linked to IFIS risk). The paper concludes that while AI shows promise, additional research is required on other imaging modalities and further model optimization to improve prediction accuracy and generalization before clinical implementation. The study does not claim that AI can currently replace clinical judgment but rather serves as a potential complementary tool for safer surgical planning.