Background
Vulvar squamous cell carcinoma accounts for 90% of vulvar cancers and is primarily treated by surgical excision. To reduce the risk of recurrence, surgeons aim to remove tumor-free margins of 2–3 mm, which can lead to extensive tissue loss, among others. Accurate intraoperative margin assessment is critical, but it currently relies on the surgeon’s experience and judgment, as well as time-consuming frozen section analysis, both of which are limited in precision.
Hyperspectral imaging offers a promising solution. This technique captures both spatial and spectral tissue information. It is a non-contact, non-invasive technique that is based on the interaction of light with the target object. However, implementation remains limited, as the current systems are restricted to image capture only.
This study aims to bridge the gap between hyperspectral imaging data and its practical use in surgery. The primary goal is to design and evaluate an interface that supports intraoperative decision-making for gynecological oncologists, ensuring the medical specialist remains in control. Since interpretability is essential, the second goal is to enhance the transparency of the HSI classification model by applying an explainable AI method.
Design Process
Research was conducted concerning the user, the context in which the user operates, existing medical interfaces, and optimal visualization of clinical data. A user survey and usability test were conducted to evaluate the prototype, which was developed through an iterative design process. The research produced a user persona, a context and task analysis, insights into medical interface layouts, and guidelines for visualizing clinical data. These aspects formed the basis for the requirements and wishes for the interface, in conjunction with the principles of the design frameworks applied. The user survey (n = 40, with a medical background) provided design direction and recommendations. The usability test was designed to evaluate the overall workload of the designed interface and its intuitiveness. The test was performed on nine participants.
The mean accuracy of all the tasks performed was 88%, indicating that the interface is straightforward to use. NASA-TLX scores ranged from 6.7 to 24 across six parameters, suggesting low to moderate cognitive load. Semantic differential results showed high ratings for user-friendliness (7.8), logical structure (8.2), and color aesthetics (7.9). However, improvements are still needed in visualizing classification outcomes and tissue parameters, which were partially addressed in the redesign.
Data Analysis
As a first step in distinguishing healthy from tumor tissue, hyperspectral data was collected intraoperatively before tumor excision. The data was then preprocessed by data calibration, glare removal, and data extraction. The preprocessed data was classified with a Support Vector Machine. This classification model was first optimized by tuning the kernel type and then validated through different evaluation metrics. To interpret the model’s decision-making, SHapley Additive exPlanations was applied to identify which features contributed most to the classification. Three types of input features were assessed: wavelengths (500–1000 nm), visual fractions, and tissue parameters.
The tuning of the kernel type yielded the Radial Basis Function kernel as the optimized model, with an average area under the curve of 1 and an accuracy of 99%. The SHapley Additive exPlanations analysis of the spectral dataset showed that all wavelengths consistently support the prediction for healthy tissue and oppose in predicting tumor. Suggesting that the model is biased toward classifying samples as healthy, and tumor is identified primarily by the absence of features associated with healthy tissue. Notably, the 500–595 nm range and the near-infrared region were most influential in supporting prediction for healthy tissue and opposing tumor.
Using spectral fractions and tissue parameters as input features did not yield accurate enough results(macro AUC = 0.65 and accuracy is 10%) to apply SHAP. This suggests that these features alone do not provide sufficient discriminatory information. This displays the need for additional spectral data to improve model performance for these input features.
Conclusion
The results of this study indicate that the initial probabilistic interface design is relatively intuitive, requiring only low to medium cognitive effort and workload. The application of explainable AI revealed that the 500–595 nm range and the near-infrared region are most influential in predicting healthy tissue versus tumor. This improves the transparency and interpretability of the classification model for gynecological oncologists. Together, the intuitive probabilistic interface and explainable AI support the integration of hyperspectral imaging into surgical practice.