DSC Europe 24 SCHEDULE

The talk provides an overview of advanced techniques for enhancing mammography diagnostics, spotlighting multiple computer vision topics. Drawing conclusions from a real-world industrial application, it delves into the technical intricacies of cutting-edge AI methodologies such as high-resolution image processing, diffusion models and anomaly detection, while also comparing these approaches to traditional ones. Through practical experiences, the talk highlights the transformative impact of these technologies on diagnostic accuracy and efficiency, underscoring their scientific contributions.

The integration of artificial intelligence (AI) with medical diagnostics is revolutionizing veterinary technology, particularly in the diagnosis of ocular conditions in dogs. This approach leverages computer vision and deep learning to significantly enhance diagnostic accuracy and efficiency. By employing sophisticated computer vision techniques, detailed analysis of canine eye imagery becomes possible, offering a new dimension in veterinary diagnostics. Additionally, Large Language Models (LLMs) are utilized to further aid the diagnostic process, showcasing the synergy between AI algorithms and veterinary medicine. The exploration of these AI-driven methods reveals their potential to streamline veterinary diagnostics, making them faster and more practical. Attendees will gain a comprehensive understanding of the technical approaches used, the challenges encountered in implementation, and the advancements achieved in applying AI to veterinary science, highlighting its transformative impact on medical diagnostics.

Over 75% of applications for clinical approval to the U.S. Food & Drug Administration (FDA) that have an AI component were submitted in the field of radiology. The number of these applications is exponentially growing, making it difficult for the FDA to process them efficiently and evaluate their performance. When it comes to evaluating new AI technologies in radiology, the following must be considered: interpretability, bias and fairness, stress testing (performance of algorithm in edge cases), intended use, etc. The “Garbage in garbage out” mantra has an emphasized importance in healthcare because of the significant hurdles to collecting and structuring medical data for AI applications. Most medical imaging data follows the strict General Data Protection Regulation (GDPR) and the new AI Act further regulates the development of the use of this data in the EU. Anatomical variability and variability of clinical conditions can be significant and the cost of acquiring large data quantities can be overwhelming. For many tasks, manual labeling is necessary which further slows down the data preprocessing speed. Labeling or any form of data annotation is challenging in medical imaging as it requires expert intervention and ground truth is generally not available. Finally, a space of features (also known as null space) that do not contribute meaningfully to the desired prediction task is abundant in medical imaging because of the high- dimensionality of data, scanner and protocol differences, presence of noise and artifacts (especially in X-ray imaging techniques where radiation dose should be kept as low as reasonably possible). Synthetical medical imaging data generation techniques have recently been considered of great potential to tackle this problem by both AI solution providers and regulatory bodies. FDA has a publicly available “Regulatory Science Tool Catalog” which contains 24 computational models (as of November 2024) for constructing some form of synthetic data for technical evaluation of new solutions. Synthetic data can be divided into stochastic (modeling patient population) and non- stochastic (individual patient modeling). Non-stochastic models can be personalized (popular “digital twins”) or model families of patents. Stochastic models can be knowledge-based (derived from physical or biological measurements) or image-based (derived from patient CT, MRI, US, or some other imaging modality). This talk will present a case study of how synthetic data can be used for product development within a startup for radiological applications. We will focus on particular knowledge-based stochastic models developed at Duke University’s Center for Virtual Imaging Trials. Then we will contrast them to image based methods that use AI tools such as Generative Adversarial Networks (GAN) and Variational Autoencoders (VAE) for statistical learning to model data distributions without requiring explicit knowledge of anatomy or physics. Finally, we will put it all in the context of the commercialization potential of solutions based on this data and contrast it to the traditional approach that relies on physical data obtained from patients in clinics.