A team at Leipzig University’s Innovation Center Computer Assisted Surgery combines hyperspectral imaging with a 3D convolutional neural network to classify tissue as healthy or malperfused. By analyzing oxygen saturation and spectral patterns across days, the system achieves an 82% AUC for early flap perfusion monitoring.
Key points
- Hyperspectral imaging captures reflectance from 540–1000 nm to compute StO₂ and NPI.
- SMOTE oversampling balances training data for rare malperfused pixels.
- A 3D CNN with 3×3 spatial patches processes spectral and perfusion inputs.
- Leave-one-patient-out cross-validation yields robust 0.82 AUC measurement.
- Model achieves 70% sensitivity and 76% specificity for flap viability.
Why it matters: Automated AI-driven monitoring of flap perfusion could revolutionize postoperative care by detecting ischemic complications earlier than clinical inspection. This approach offers non-invasive, objective assessments, potentially improving flap salvage rates and reducing surgical revision.
Q&A
- What is hyperspectral imaging?
- How does a convolutional neural network analyze perfusion data?
- What are NPI and StO₂ metrics?
- Why use SMOTE oversampling?
- What is flap malperfusion?
