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

Deep Anterior Lamellar Keratoplasty (DALK) is a partial-thickness corneal transplant procedure for treating corneal stromal diseases. A key step involves 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 proposed an Optical Coherence Tomography (OCT)-guided, eye-mountable robot that utilizes real-time corneal layers tracking from M-mode OCT signals for control. However, signal noise and instability during the manipulation of the OCT fiber sensor-integrated needle hinder the performance of conventional deep-learning segmentation methods, leading to rough and inaccurate detection of corneal layers. To overcome these limitations, we develop a topology-based deep-learning segmentation method that integrates a topological loss function with a modified network architecture. This approach effectively mitigates noise effects and enhances segmentation precision, stability, and speed. Validation on in vivo, ex vivo, and hybrid rabbit eyes data sets demonstrates that our approach surpasses traditional loss-based techniques, delivering accurate, robust, and fast segmentation of the epithelium and DM for surgical guidance.