Adaptive Sensitivity-Fisher Regularization Enhances Vascular Segmentation in Laparoscopy
Overview
This study introduces an Adaptive Sensitivity-Fisher Regularization (ASFR) method within a heterogeneous transfer learning framework to improve vascular segmentation in laparoscopic videos using only white light imaging. The approach addresses challenges of limited annotated data and domain differences, achieving robust segmentation performance across advanced video segmentation architectures.
Background
Accurate vascular localization during laparoscopic surgery is critical to avoid complications such as bleeding, especially in procedures like laparoscopic sleeve gastrectomy. Traditional methods rely on indocyanine green (ICG) fluorescence or Doppler ultrasound, which have workflow and safety limitations. Continuous vascular segmentation using only white light imaging could simplify intraoperative workflows and enhance patient safety. However, the scarcity of annotated laparoscopic vascular datasets hinders deep learning model development. Heterogeneous transfer learning offers a solution by transferring knowledge from large non-medical datasets to the medical domain, but adapting complex video segmentation networks remains challenging due to risks of catastrophic forgetting and negative transfer.
Data Highlights
The proposed ASFR method integrates Fisher information and sensitivity analysis to identify critical parameters for fine-tuning pretrained video segmentation networks such as STCN and XMem. This approach preserves essential pretrained knowledge while optimizing adaptation to laparoscopic vascular segmentation. Training uses annotations from the first video frame to segment vessels in subsequent frames, including non-consecutive frames, enabling comprehensive temporal evaluation.
Key Findings
ASFR effectively mitigates catastrophic forgetting by regularizing parameters based on their Fisher information and sensitivity during fine-tuning.
Heterogeneous transfer learning enables leveraging large-scale non-medical video datasets to improve vascular segmentation despite limited annotated laparoscopic data.
The method achieves robust and continuous segmentation of both visible and obscured vessels using only conventional white light laparoscopic videos.
ASFR enhances performance across diverse video segmentation architectures, demonstrating broad applicability.
Continuous vascular localization can potentially reduce intraoperative bleeding risks by improving vessel identification before ligation.
Clinical Implications
The ASFR-based transfer learning framework allows surgeons to rely solely on white light imaging for continuous vascular localization, eliminating the need for additional imaging modalities like ICG fluorescence. This simplification can reduce procedural complexity and associated risks such as allergic reactions. Improved vascular segmentation accuracy may enhance intraoperative decision-making and patient safety during laparoscopic procedures.
Conclusion
The study presents a novel ASFR method within a heterogeneous transfer learning framework that successfully adapts pretrained video segmentation models for accurate vascular localization in laparoscopic videos using white light imaging. This approach addresses key clinical challenges and advances the potential for safer, more efficient laparoscopic surgeries.
