Topology-based deep-learning segmentation method for deep anterior lamellar keratoplasty (DALK) surgical guidance using M-mode OCT data

TL;DR

The paper tackles real-time segmentation of corneal layers from noisy M-mode OCT signals to guide DALK surgery. It introduces a topology-based loss with a star-shaped prior integrated into a modified U-Net, yielding a hybrid loss $L_{hybrid}$ that balances pixel accuracy with geometric consistency $L_T$ alongside standard $L_{BCE}$. Evaluation across in vivo, ex vivo, and hybrid rabbit datasets shows the proposed method outperforms conventional loss-based segmentation in SSIM, PSNR, IoU, and Dice metrics, while achieving higher inference speeds (up to $40$ Hz) and more precise layer tracking for epithelium and Descemet's membrane. The approach demonstrates robustness to noise and motion artifacts, with direct implications for improved, real-time robotic guidance during DALK procedures.

Abstract

Deep Anterior Lamellar Keratoplasty (DALK) is a partial-thickness corneal transplant procedure used to treat corneal stromal diseases. A crucial step in this procedure is the precise separation of the deep stroma from Descemet's membrane (DM) using the Big Bubble technique. To simplify the tasks of needle insertion and pneumo-dissection in this technique, we previously developed an Optical Coherence Tomography (OCT)-guided, eye-mountable robot that uses real-time tracking of corneal layers from M-mode OCT signals for control. However, signal noise and instability during manipulation of the OCT fiber sensor-integrated needle have hindered the performance of conventional deep-learning segmentation methods, resulting in rough and inaccurate detection of corneal layers. To address these challenges, we have developed a topology-based deep-learning segmentation method that integrates a topological loss function with a modified network architecture. This approach effectively reduces the effects of noise and improves segmentation speed, precision, and stability. Validation using in vivo, ex vivo, and hybrid rabbit eye datasets demonstrates that our method outperforms traditional loss-based techniques, providing fast, accurate, and robust segmentation of the epithelium and DM to guide surgery.

Figures (5)

• Figure 1: Example of rough and inaccurate segmentation caused by signal noise and fluctuations. Blue Box: Hollow regions and irregular boundaries resulting from information loss and noise interference.
• Figure 2: (a) Geometric illustration of the star shape prior in the topological loss. (b) Modified U-Net architecture.
• Figure 3: Comparison of segmentation results between the proposed method (PM) and the conventional method (CM). (a) Representative example from the in vivo dataset. (b) Representative example from the ex vivo dataset.
• Figure 4: Evaluation pipeline for the average absolute pixel error in corneal layer segmentation. $E_i$ and $E_i^\prime$, as well as $D_i$ and $D_i^\prime$, denote the vertical pixel positions of the epithelium and Descemet's membrane for the ground truth and predicted result, respectively.
• Figure 5: Video 1-Sample video of real-time segmentation and tracking of corneal layers using the proposed method. http://dx.doi.org/doi.number.goes.here