Real-time topology-aware M-mode OCT segmentation for robotic deep anterior lamellar keratoplasty (DALK) guidance

Robotic deep anterior lamellar keratoplasty (DALK) requires accurate real time depth feedback to approach Descemet’s membrane (DM) without perforation. M-mode intraoperative optical coherence tomography (OCT) provides high temporal resolution depth traces, but speckle noise, attenuation, and instrument induced shadowing often result in discontinuous or ambiguous layer interfaces that challenge anatomically consistent segmentation at deployment frame rates. We present a lightweight, topology aware M-mode segmentation pipeline based on UNeXt that incorporates anatomical topology regularization to stabilize boundary continuity and layer ordering under low signal to noise ratio conditions. The proposed system achieves end to end throughput exceeding 80 Hz measured over the complete preprocessing inference overlay pipeline on a single GPU, demonstrating practical real time guidance beyond model only timing. This operating margin provides temporal headroom to reject low quality or dropout frames while maintaining a stable effective depth update rate. Evaluation on a standard rabbit eye M-mode dataset using an established baseline protocol shows improved qualitative boundary stability compared with topology agnostic controls, while preserving deployable real time performance.

💡 Research Summary

This paper addresses a critical need in robotic deep anterior lamellar keratoplasty (DALK): providing surgeons with reliable, real‑time depth feedback to approach Descemet’s membrane (DM) without perforation. While intra‑operative M‑mode optical coherence tomography (OCT) offers high‑temporal‑resolution depth traces at a fixed lateral position, the raw signals are plagued by speckle noise, signal attenuation, and instrument‑induced shadowing. These degradations frequently produce fragmented or ambiguous layer interfaces, causing conventional pixel‑wise segmentation networks to output discontinuous or anatomically implausible boundaries when operating at deployment frame rates.

The authors propose a lightweight, topology‑aware segmentation pipeline built on the MLP‑based UNeXt architecture. UNeXt processes single‑channel intensity inputs and outputs two classes (epithelium and DM). To satisfy down‑sampling constraints, inputs are zero‑padded to multiples of 16, processed, and then cropped back to the original size. Automatic mixed precision (AMP) is employed during inference, and the M‑mode frames are tiled into eight non‑overlapping vertical strips (512 × 64) to maximize GPU batch throughput. After inference, the strips are reassembled for visualization and metric computation.

The key novelty lies in the loss formulation. The training objective combines binary cross‑entropy (BCE), Dice loss, and a topology‑aware regularization term weighted by λ_topo. The topology term implements a star‑shape prior that penalizes fragmented boundaries and anatomically impossible configurations along the depth axis. λ_topo is ramped up gradually during training to avoid early over‑regularization, allowing the network to first learn accurate pixel‑wise predictions before enforcing global anatomical consistency.

The dataset consists of rabbit‑eye M‑mode OCT images widely used for DALK research: 500 in‑vivo and 250 ex‑vivo annotated frames (512 × 512). A hybrid set merges both, with train/test splits of 400/100 (in‑vivo), 200/50 (ex‑vivo), and 600/150 (hybrid). Two interfaces are annotated: the upper epithelium boundary and the lower DM boundary. Pixel spacing is 2.61 µm/px. All distance‑based metrics are reported in micrometers.

Performance is evaluated on several fronts:

1. Segmentation quality – Macro Dice/IoU, SSIM, PSNR, and mean absolute boundary error (both in pixels and µm) are reported. Compared with a topology‑aware U‑Net baseline, the proposed UNeXt (topology‑aware) achieves comparable Dice (0.9332 vs 0.9874) but markedly higher PSNR (31.83 dB vs 29.41 dB) and SSIM (0.9801 vs 0.9901). More importantly, average boundary error drops from 0.34 px (0.89 µm) for epithelium and 1.26 px (3.29 µm) for DM (U‑Net) to 0.31 px (0.98 µm) and 0.55 px (1.70 µm) respectively for UNeXt, indicating substantially more stable tracking under low‑SNR conditions.

2. Real‑time throughput – The authors measure end‑to‑end frame rate across the full pipeline (pre‑processing, host‑device transfer, batched GPU inference, strip reassembly, post‑processing, and solid‑band overlay rendering). The system sustains >80 Hz on a single GPU, exceeding typical intra‑operative display rates (30–60 Hz) and providing temporal headroom for frame‑quality gating. This redundancy enables the rejection of low‑quality frames without compromising the effective update rate, a crucial safety feature for robotic control loops.

3. Robustness analysis – Temporal robustness is examined by simulating speckle bursts, attenuation spikes, and shadowing events. The high frame rate allows a lightweight confidence‑based gating mechanism that discards corrupted frames while maintaining a stable visual feed for the surgeon and the robot controller.

The discussion acknowledges limitations: reliance on a single animal dataset, and the inability of topology regularization alone to recover completely lost interfaces. Future work will explore confidence‑aware gating, closed‑loop safety constraints, and validation across multiple OCT platforms and human clinical scenarios.

In conclusion, the paper demonstrates that integrating a fast MLP‑based segmentation backbone (UNeXt) with a star‑shape topology prior yields a practical, real‑time solution for M‑mode OCT segmentation in robotic DALK. The approach bridges the gap between high‑accuracy anatomical segmentation and the stringent latency requirements of intra‑operative guidance, offering a promising pathway toward autonomous or semi‑autonomous corneal surgery.