Article
Intelligent ultrasonic-aspirator for CNS/ tumor tissue differentiation – a feasibility study using machine learning
Intelligenter Ultraschall-Aspirator zur Differenzierung von ZNS/Tumor Gewebe: Eine Machbarkeitsstudie mit maschinellem Lernen
Search Medline for
Authors
-
Niclas Bockelmann
- Universität zu Lübeck, Institut für Robotik und Kognitive Systeme, Lübeck, Deutschland
-
Daniel Schetelig
- Söring GmbH, Quickborn, Deutschland
-
Mario Matteo Bonsanto
- Universitätsklinikum Schleswig-Holstein, Klinik für Neurochirurgie, Lübeck, Deutschland
-
Steffen Buschschlüter
- Söring GmbH, Quickborn, Deutschland
-
Floris Ernst
- Universität zu Lübeck, Institut für Robotik und Kognitive Systeme, Lübeck, Deutschland
| Published: | May 25, 2022 |
|---|
Outline
Text
Objective: The goal of brain tumor surgery is the complete removal of all tumorous tissue while not damaging healthy, tumor-free brain areas. Therefore, emphasis must be placed on the intraoperative differentiation of brain tumors and healthy brain tissue. Studies have shown that healthy brain tissues differ from tumor tissues in their mechanical properties. One commonly used instrument for tumor removal is the ultrasonic-aspirator, which is responsive to tissue properties. In in this feasibility study, an approach for tissue differentiation based on this electrical response using machine learning will be investigated.
Methods: Three synthetic tissue models, different in their mechanical properties (i.e. Young's modulus: soft, medium and firm), are used and represent healthy (soft) and tumorous tissue (medium/firm). For each tissue model, two generator settings (low and high) are applied. More than 38.000 temporal measurement points are recorded during tissue ablation for each of the electrical signals produced by the ultrasonic-aspirator. Training and test splits are defined, resulting in 87.5% and 12.5% of the data respectively. Classification is done with two machine learning approaches: an AdaBoost (AB) and a neural network (NN) classifier. Since the settings of the generator are known, they are used as prior information. This results in two setting-based classifiers, in addition to one general single classification model over all generator settings.
Results: Results are given in the metrics accuracy, F1-score, precision and recall. The results can be seen for the general single classifier and ultrasound-based classifiers in Table 1 [Tab. 1] demonstrating a good performance regarding differentiation of mechanical properties. Furthermore, the results show that the usage of prior information of the generator settings can improve the results regardless of the classification method.
Conclusion: The results indicate the feasibility to differentiate tissue properties, like soft, medium and firm, using a commercial ultrasonic-aspirator in a laboratory environment. Future work needs to focus on a more diverse dataset in terms of tissue properties and generator settings to capture the variety present in a real intraoperative setting.
