Publications

Abstract : Biomedical brain imaging lies at the interface of visual, and spatial neuropathology and neurosurgical intervention. Future treatments are expected to require greater feature detail in imaging as well as underlying mechanisms that will be synergistically advanced. This will be made more accessible via low power embedded devices and cloud platforms through applied deep learning software and hardware analysis. Towards that end, we develop a hardware-software co-design technique referred to as SCENIC (or, Separable Convolution Enabled Non-Invasive Classification) for the identification and classification of glioma brain tumors, using physical tissue features reflected in parameter weighted MRI scan types such as-T1-w, T1-ce, T2-w and FLAIR. The high performance hardware exceeds current accuracy, resource efficiency and time consumption parameters. The proposed SCENIC-CNN Accelerator is synthesized on 45 nm process technology andit can operate at a minimum frequency of 1GHz while maintaining low-power consumption of only 0.36 W and a low chip-area size of 0.431mm2. Our classification accuracy achieves 98.3% in detection of the presence of a tumor pathology and 99.62% within classification of imaging modalities that relate to tissue parameters such as fat content, blood or CSF flow and tissue density. Compared to prominent state-of the-art Convolutional Neural Network (CN) models being designed for biomedical imaging, SCENIC is competitive versus XceptionNet, InceptionV3, ResNet-50, and VGG-16. With model compression techniques, SCENIC requires a memory space of less than 0.265MB. We discuss design methodology as applied to future goals to meet challenging needs to distinguish tumor origins, such as glioma and metastasized tumors, along with other neuropathologies that may be TBI, vascular or developmental. SCENIC will aid impactful, cost-effective, rapid and accurate neurosurgical intervention and treatments.