TL;DR
This paper introduces a novel, non-invasive sensor system that automatically detects electrosurgical cautery states (like "cut" vs. "coagulate" and energy levels) during surgery. This advancement promises to enhance safety and precision in computer-assisted and robotic surgeries by providing real-time feedback on tool-tissue interaction.
What Problem Does This Paper Address?
In computer-assisted surgery (CAS) and robotic surgery, accurately knowing when a surgical tool is interacting with tissue is crucial. This "energy event" is particularly important when using electrosurgical cautery, a common tool for cutting and sealing tissue. Current methods often rely on manual surgeon input or complex, invasive systems. The challenge is to automatically and reliably detect these energy events and the specific settings of the cautery device (like power level and mode) without disrupting the surgical workflow or requiring modifications to existing, approved surgical equipment. This information is vital for precise tissue classification, improving surgical outcomes, and potentially reducing complications like positive margins in cancer surgeries.
Methods
The researchers developed a sensor-based system to capture electrical signals from the cautery device. They used non-invasive current sensors and an oscilloscope to record voltage changes in the cautery and grounding pad wires. This data was then processed using an open-source software module integrated with 3D Slicer, a platform commonly used in surgical navigation.
The core of their approach involves machine learning:
- Signal Processing: Fast Fourier Transform (FFT) was applied to the voltage signals to analyze their frequency components.
- Feature Extraction: Principal Component Analysis (PCA) was used to reduce the dimensionality of the FFT data and identify key features. Alternatively, they extracted the maximal intensity and corresponding frequency from the FFT output.
- Classification: Two machine learning models, Support Vector Machine (SVM) and Random Forest Classifier (RFC), were trained to classify the different cautery states (off, cut-air, cut-tissue, coagulate-air, coagulate-tissue) and energy levels.
- Workflow Integration: The system was designed to be non-disruptive, attaching to existing tools and feeding data into the surgical navigation software in real-time.
Datasets
The study utilized ex vivo tissue samples (chicken, porcine, and bovine) to simulate surgical conditions. They tested two different electrosurgical generator units (Valley Lab Force FX C and ConMed System 5000) at three power levels (30W, 35W, 40W) in both "cut" and "coagulate" modes. Data was collected for 120 incisions per test setup, with each incision lasting approximately 1 second. The data was split into training and validation sets, with an 8-fold cross-validation strategy employed for evaluation.
Key Insights and Ideas
- Non-invasive and Workflow-Friendly: The system's ability to detect cautery states using external sensors without altering surgical tools or workflow is a significant advantage for clinical adoption.
- High Accuracy Potential: The sensor-based approach, coupled with machine learning, demonstrated a remarkable average accuracy of 95.56% in classifying cautery states across various tissue types and energy settings.
- Distinguishing Subtle States: The system successfully differentiated between subtle states like "cut-air" and "cut-tissue," which is critical for pinpointing the exact moment of tissue interaction.
- Energy Level Detection: Beyond just detecting the energy event, the system can also identify the specific power level being used, offering more granular control and information.
- Open-Source Approach: Providing the software as open-source is a crucial step for broader research and development, enabling replication and further innovation in the field.
Key Results
The Random Forest Classifier (RFC) consistently outperformed the Support Vector Machine (SVM) in most scenarios.
- Cautery State Classification: Achieved an average accuracy of 98.52% for tissue type, 97.98% for power level, and 86.00% for cautery device when using principal components. When using maximal intensity and corresponding frequency, accuracies were even higher, reaching 98.93% for tissue and 97.22% for power.
- Energy Level Detection: The RFC achieved accuracies of 99.05% for ConMed and 99.93% for Valley Lab devices when using maximal intensity and corresponding frequency.
- The models performed best when classifying based on tissue type, followed by power level. Classification was more challenging when distinguishing between cautery devices, likely due to differences in their electrical signal characteristics.
Limitations and Future Work
- Limited Sample Size: The study acknowledges a small sample size, and the use of animal tissue rather than human tissue might limit the direct applicability to clinical scenarios. Future work should include in vivo testing with human tissue.
- ESU Variability: The models showed reduced performance when one cautery device was excluded from training, highlighting the need for separate models for different electrosurgical units (ESUs) due to variations in their frequency signatures.
- Grounding Pad Location: The grounding pad location was not varied during experiments, and its position could influence the current signal. Future studies should investigate this effect.
- Interference from Other Tools: The potential for interference from other electrosurgical tools used in the operating room was not evaluated.
- Spatial Accuracy: The current study focused on temporal detection of cautery states. Future work should integrate this system with spatial tracking (e.g., EM tracking) to evaluate the accuracy of locating surgical incisions.
- Model Robustness: Investigating the impact of viscoelasticity on tissue classification and exploring combinations of feature extraction methods (PCA and maximal intensity/frequency) could further enhance model robustness.
- Clinical Validation: The ultimate goal is to collect data from actual breast cancer surgeries to validate the model's performance in a real clinical setting.
Summary updated