|
Tissue classification in surgical workflows between healthy and tumoral regions remains challenging both during and post-surgery. The current standard practice consists of taking small biopsies directly after tumor resection and sending them to pathologists for an intraoperative margin assessment, which is time-consuming and error prone due to the necessarily limited size and number of samples. Then, after the surgery is completed, the resected tumor is sent to the pathology lab, where its type and grading are further confirmed. The present workflow is prone to inaccuracies and particularly difficult when the sample is resected in several pieces. Therefore, an intraoperative tissue classification technology is highly sought-after for a simplified surgical workflow and better patient outcome. Our work aims at using hyperspectral images (HSI) for contact- and tracer-free tissue differentiation. We introduce a deep learning-based algorithm for the classification of tissue type that is based on spectral information and can be applied simultaneously to the whole sample. We illustrate the performance of our method on ex vivo head and neck squamous cell cancer samples. The proposed algorithm can differentiate between three main classes: background, tumor, and healthy tissues. Our experiments first assess the generalization of the neural network on data from unseen cases. We then determine the minimal number of training examples needed to cover the variety of tissue spectral appearances seen in the clinical dataset. We evaluate the influence of the delay between resection and start of image acquisition on the quality of the recorded HSI and the prediction. Qualitative and quantitative evaluations support the applicability of hyperspectral imaging for tissue classification and demonstrate an agreement between surgeon annotations and neural network predictions in most test cases.
|
