Human glioma infiltration detection algorithm for Optical Coherence Tomography: an AI-Assisted approach based on tissue simulating phantoms

Gliomas are primary brain tumors with a high rate of malignancy: they account for 28% of all brain tumors and 80% of malignant brain tumors. Maximal tumor resection during surgery is one of the main goals in the treatment of these cancers as it improves the quality of life of the patients and their survival rate. In order to achieve maximal resection, glioma-infiltrated tumor margins must be correctly detected by a system that works in real time. To fulfill this purpose, traditional Magnetic Resonance Imaging (MRI) presents several limits, like elevated costs and a bulky device, and the development of in situ, fast segmentation algorithm is necessary. The use of Optical Coherence Tomography (OCT) is spreading to a larger range of applications other than ophthalmology and dermatology, due to its non-invasiveness, high resolution, and imaging speed. These characteristics have led this technique to be often associated with segmentation tools in the diagnosis and investigation of several pathologies, even brain cancer. In this Thesis work, an Artificial Neural Network (ANN) for the detection of human Glioma infiltration in OCT images is presented. The algorithm is trained on acquisitions performed on human tissue-simulating phantoms made of silicone, that can mimic both healthy brain tissue and glioma-infiltrated tissue, and then tested on real tissue acquisitions obtained from brain biopsies. Tissue phantoms are widely used in both clinic and research as they provide a powerful tool for device testing and performances evaluation, without the need of real tissue samples, and they can be easily reproduced. The training and validation phases are carried out on a large dataset of approximately 2 million samples (A-lines) for an extensive amount of time (~500 epochs). Unlabeled OCT images from 4 biopsies of High-grade Gliomas have been used as test set to establish the detection performances: overall results have demonstrated a sensitivity of ~94% and a specificity of ~85%, but higher values can be achieved considering that most of errors come from small defects within the tissue sample that are not present in an intraoperative situation. This algorithm can be an efficient and useful instrument as OCT-guided surgical tool, providing to the surgeon a 3D representation and classification of the tissue. Further developments will include the reduction of software’s computational time, possibly with the use of a dedicated GPU, and a tuning of the algorithm for different wavelengths of the OCT, to create a more general system that can be interfaced to machines with different characteristics.