Neuromorphic Medical Image Analysis at the Edge : On-Edge training with the Akida Brainchip

Abstract: Computed Tomography (CT) scans play a crucial role in medical imaging, allowing neuroscientists to identify intracranial pathologies such as haemorrhages and malignant tumours in the brain. This thesis explores the potential of deep learning models as an aid in intracranial pathology detection through medical imaging. By first creating a convolutional neural network model capable of identifying brain haemorrhage and then moving it onto the neuromorphic processor Akida AKD1000, it allowed the usage of Spiking Neural Networks and on-edge retraining capabilities. In a process called few-shot learning, the model was trained to also identify brain tumours with minimal additional samples. The research further investigated how the parameters used in the edge-learning influenced classification accuracy. It was shown that the parameter selection and interaction introduced a trade-off in regard to accuracy for the haemorrhage and tumour classification models, but an optimal constellation of parameters could be extracted. These results aim to serve as a foundation for future endeavours in image analysis using neuromorphic hardware, specifically within the domain of few-shot and on-edge learning. The integration of these models in the medical field has the potential to streamline the diagnosis of intracranial pathologies, enhancing accuracy and efficiency while unloading medical professionals.