Passer au contenu de la page principale

Using Deep Learning for Phase Recognition and Training Metrics in Cataract Surgery - 5606

Mon statut pour la session

Quand:
4:45 PM, Vendredi 20 Juin 2025 (5 minutes)
Author’s Name(s): Joshua Bierbrier, Rebecca Hisey, Adrienne Duimering, Gabor Fichtinger, Matthew Holden, Christine Law

Author’s Disclosure Block: Joshua Bierbrier: none; Rebecca Hisey: none; Adrienne Duimering: none; Gabor Fichtinger: none; Matthew Holden: none; Christine Law: none

Abstract Body
Purpose:Phase recognition in cataract surgery can introduce quantitative metrics, like phase completion times, in resident surgical training programs. Resident training progression can be objectively quantified as proficiency increases. Phase completion times can additionally be used to optimize operating time allocation for residents. Deep learning models have been validated for phase prediction in several surgical procedures. We explore the use of deep learning-based phase recognition to assess phase completion times in cataract surgery and develop objective metrics for trainee feedback. Study Design: Microscope-recorded videos of uncomplicated cataract surgeries were collected from Kingston Health Sciences Centres and used to train a deep learning model. The model was evaluated using a cross-validation scheme. Methods: The dataset consists of 30 videos (15 staff, 15 residents). Twelve ground truth phases were manually annotated by an expert for each video. We trained a deep learning model to predict the phase of surgery from video frames. A pre-trained ResNet50 model first extracts spatial features from each frame. A Long Short-Term Memory model then examines these features across consecutive frames to predict the phase for each frame.The model was trained and evaluated using 5-fold nested cross-validation. To assess the performance of the model, accuracy, F-score (a measure of precision and recall), and Jaccard Index (a measure of overlap) were calculated. The Mann-Whitney U test with Holm-Sidak corrections was used to assess significant differences in phase times between staff and residents, evaluating the model's ability to detect these differences. Results: The model achieves an accuracy of 83.3%, F-score of 0.83, and Jaccard Index of 0.75 on the test data. Based on the per-phase F-score, the model performed best on phacoemulsification (0.95), capsulorhexis (0.89), viscoelastic removal (0.84), and lens insertion (0.83). The model performed worst on hydration (0.50) and lens positioning (0.49). The model’s phase predictions identify a statistically significant difference between staff (213.8±57.3s) and residents (417.3±129.0s) to complete phacoemulsification (p<0.05), which agrees with the ground truth data (Staff:230.7±57.9s; Residents:438.1±166.8s). Conclusions: The results are promising for the development of objective training progression metrics. The model performs particularly well recognizing phacoemulsification, which is considered the most challenging phase for residents to learn. By detecting differences in phase completion times between staff and residents that align with ground truth data, the model demonstrates its potential as a valuable tool for quantifying trainee performance. Increased training data can improve the model’s performance on phases like hydration and lens positioning.

Joshua Bierbrier

Conférencier.ère

Mon statut pour la session

Évaluer

Detail de session
Pour chaque session, permet aux participants d'écrire un court texte de feedback qui sera envoyé à l'organisateur. Ce texte n'est pas envoyé aux présentateurs.
Afin de respecter les règles de gestion des données privées, cette option affiche uniquement les profils des personnes qui ont accepté de partager leur profil publiquement.

Les changements ici affecteront toutes les pages de détails des sessions