ECCV 2020 Daily - Wednesday

Challenges of medical image analysis Computer vision techniques have been used in medical imaging for many tasks including detection, classification, prediction, noise reduction, image registration and automation of many tedious manual tasks! In his second talk Keith looks at some Challenges and Practical Pitfalls of working with medical images. Convolutional Neural Networks (CNNs) have become the tool of choice for the majority of image analysis tasks, and two of the biggest challenges facing deep learning in medical image analysis are access to data and generation of ground truth labels. In general, both are harder for medical imaging than other imaging applications due to privacy concerns surrounding personal health records and the high level of specialist expertise that is often required to annotate medical images. We firmly believe that, used correctly, data saves lives – but by by extension, lack of data due to unnecessarily restrictive controls could potentially cost lives. It is vital that the privacy versus public benefit balance is correct. Keith’s biggest practical tip is to always look at your data, scrutinise what your model is seeing, and don’t just rely on summary metrics. Deploying medical AI in the real world 3 CV4MI 21 DAILY W e d n e s d a y 5 Medical image analysis in action! Vessel segmentation & tracking followed by CT-X-Ray image registration for fluoroscopy-guided intervention, to support transarterial chemoembolization (TACE) for liver cancer [Embolization Plan&Multi Modality Roadmap applications, CanonMedical]. In this talk, Alison discusses 4 real-world case studies of medical image analysis, spanning a range of clinical and vision tasks. AI is offering game-changing solutions for some applications such as image reconstruction (denoising). For other tasks such as vessel tracking and image registration (see Fig. 5 ), we continue to leverage classical techniques, often in combination with deep learning for the recognition steps.