Introduction: Accurate intraoperative tumour margin assessment in vulvar cancer remains challenging. Inadequate resection margins increase the risk of local recurrence, whereas excessive removal of healthy tissue may impair function and quality of life. We evaluated whether intraoperative hyperspectral imaging (HSI) combined with deep learning can accurately distinguish tumour from healthy tissue to support surgical margin assessment.
Methods: We developed a standardised pipeline for intraoperative HSI acquisition, preprocessing and tissue classification. Reference annotations combined imaging-based labels from gynaecological oncologists with histopathological labels from pathologists. A patch-based 3D convolutional neural network was trained and evaluated in two settings: (1) training and testing on in vivo data and (2) training on ex vivo data with testing on in vivo data to assess model generalisability. Super-resolution reconstruction was applied to enhance the spatial resolution of HSI while preserving spectral information.
Results and Discussion: A total of 123 participants were included: 77 patients with vulvar cancer and 46 healthy volunteers. The model trained on in vivo data achieved a sensitivity of 91.6%, specificity of 94.9% and precision of 77.9% for tumour classification. Models trained on ex vivo data generalised to in vivo images with a high sensitivity (99.1%), but lower specificity (85.9%) and precision (48.3%), highlighting the challenges of domain shift between ex vivo and in vivo imaging. Super-resolution substantially improved image quality (SSIM from 0.613 to 0.931) while preserving spectral fidelity (SAM 0.0136) and maintaining classification performance, supporting its use for improved intraoperative visualisation.
Conclusions: Intraoperative HSI combined with deep learning enables accurate tumour classification and shows promise for improving surgical margin assessment in vulvar cancer. Further validation in larger and more diverse patient cohorts is warranted to support clinical implementation.