EdgeAI & Medical Image Segmentation

In the modern age, artificial intelligence (AI) finds diverse applications, enhancing the quality of our lives. Within this context, we have two speakers who will shed light on the utilization of AI in medical imaging and edge computing, offering insights into these innovative domains.

Towards universal medical image segmentation by Dr. Matthieu Ruthven

Image segmentation is playing an increasingly important role in healthcare, with applications ranging from heart function assessment to cancer treatment. Currently, most state-of-the-art automatic segmentation methods are AI-based. However, usually these methods are designed for a specific segmentation task and are unable to generalise to other segmentation tasks involving other body organs, types of medical images or segmentation classes. To adapt them to new segmentation tasks, these methods must be either be trained from scratch or fine-tuned, which is time-consuming and poses a substantial barrier for clinical researchers who often lack the resources and expertise to do this. In this talk, I will present UniverSeg, a recently developed and publicly available medical image segmentation method that begins to address this barrier. UniverSeg is AI-based but generalises to new segmentation tasks without requiring additional training, thus facilitating the investigation and identification of new tasks to ultimately improve the diagnosis and treatment of disease.

Hardware-Aware Neural Architecture Search Using Genetic Algorithm by Nilotpal Sinha

In the realm of artificial intelligence, edge devices play a pivotal role by bringing computational power closer to the source of data, reducing latency and enabling real-time decision-making. These devices, ranging from sensors to smartphones, are crucial for enhancing AI applications, ensuring efficiency, and unlocking the potential for decentralized and responsive systems. In this talk, I will be presenting my work on using genetic algorithm for designing the architecture of neural networks for different target edge devices. Such algorithms are called Hardware-aware Neural Architecture Search (HW-NAS). This research opens up the door to tailoring neural network architectures to specific edge devices, leading to faster and more resource-efficient AI models. This not only enhances the performance of AI applications but also enables the deployment of sophisticated algorithms in decentralized environments.