Article Impact Level: HIGH Data Quality: STRONG Summary of Scientific Reports, 15(1), 16886. https://doi.org/10.1038/s41598-025-00303-z Dr. Simon Mueller et al.
Points
- This study evaluated the performance of deep learning using the MS-TCN++ model to recognize surgical phases in Manual Small-Incision Cataract Surgery using the newly introduced SICS-105 dataset.
- The model achieved 85.56 percent accuracy and 98.22 percent ROC AUC on SICS-105, slightly lower than its performance on the existing Cataract-101 dataset.
- Phase recognition in SICS proved more complex due to its longer duration and higher number of distinct procedural steps than phacoemulsification.
- ROC AUC values for individual SICS phases varied widely, indicating the model’s inconsistent performance and the need for further refinement across diverse patient cases.
- By making the SICS-105 dataset open access, this research lays the groundwork for enhanced training, analysis, and future automation in cataract surgery, especially in low-resource settings.
Summary
This prospective cross-sectional study investigates the use of deep learning (DL) for phase recognition in Manual Small-Incision Cataract Surgery (SICS), a widely used procedure in low- and middle-income countries (LMICs). The study introduces the SICS-105 dataset, comprising 105 patients recruited from Sankara Eye Hospital in India. Using the MS-TCN++ architecture, the performance of DL phase recognition was evaluated against the well-established Cataract-101 dataset. The results show that the MS-TCN++ architecture achieves an accuracy of 85.56% [CI 80.63–92.09%] on the SICS-105 dataset, compared to 89.97% [CI 86.69–93.46%] on Cataract-101, with the ROC AUC for SICS-105 at 98.22% [97.16–99.26%] compared to 99.10% [98.34–99.51%] for Cataract-101.
While DL-based phase recognition proved effective in SICS, challenges arose due to the larger number of phases and longer procedure duration, making it more difficult than phacoemulsification, which has fewer phases. Despite these challenges, the MS-TCN++ model was able to capture phase transitions effectively, though the accuracy distribution and confidence intervals overlapped. The ROC AUC values for SICS ranged from 46.20% to 94.18%, highlighting the variability in model performance across different patients and phases. This suggests that further model refinement is needed to achieve consistent and accurate phase detection in SICS.
This research is a significant step toward improving postoperative analysis and training for SICS by incorporating DL-based phase recognition. The dataset used in this study, SICS-105, is made open access to support further developments in the field. This could potentially lead to better patient outcomes in cataract surgery and reduce the gap between manual and computer-assisted techniques.
Link to the article: https://www.nature.com/articles/s41598-025-00303-z
References
Mueller, S., Sachdeva, B., Prasad, S. N., Lechtenboehmer, R., Holz, F. G., Finger, R. P., Murali, K., Jain, M., Wintergerst, M. W. M., & Schultz, T. (2025). Phase recognition in manual Small-Incision cataract surgery with MS-TCN + + on the novel SICS-105 dataset. Scientific Reports, 15(1), 16886. https://doi.org/10.1038/s41598-025-00303-z
