Event Date
Event Date
This study presents a novel deep learning approach for real-time brain cancer detection using optical coherence tomography (OCT) images during neurosurgical procedures. We combine a pretrained Vision Transformer (DINOv2) with convolutional neural networks processing Grey Level Co-occurrence Matrix texture features in a dual-path architecture. An adversarial discriminator network mitigates patient-specific bias from limited cohort data. Using OCT images from 11 patients (7,441 B-frame slices), our model achieved 98.6% accuracy on held-out test patients not seen during training, with 3.6ms inference time, demonstrating significant potential for intraoperative decision-making in brain cancer surgeries with real-time, high-accuracy tissue classification.
Presenter
Soumyajit Ray
Johns Hopkins Univ. (United States)
Soumyajit received his B.Tech in Instrumentation Engineering from Indian Institute of Technology, Kharagpur. After completing his MBA from IIM Ahmedabad, he worked in a number of multinational companies in technology leadership roles. He recently received his MS in Biomedical Engineering from Washington University where he developed a peripheral nerve interface to reanimate paralyzed limbs using signals from an electrocorticography based brain-computer interface (BCI). He has also worked in the NeuRonICS lab at Indian Institute of Sciences, where he developed an EEG-BCI based communication device for patients with advanced ALS. He is currently working on developing a functional guidance system for prostatectomy using ultrasound stimulation, fluorescence microscopy, and photoacoustic imaging.