Suggestions
Anton Schwaighofer
Machine Learning Engineer and Lead at Microsoft Research
Anton Schwaighofer is a machine learning engineer and team lead specializing in AI for Medical Imaging. With over a decade of experience, he has a strong track record of delivering successful machine learning solutions across various domains. He holds a PhD in machine learning and a degree in Computer Science, combining theoretical expertise with practical industry experience. Anton focuses on creating accurate, well-engineered, and sustainable machine learning solutions. His technical skills include Python, .NET (F#, C#), Pytorch, and Azure stack (Datalake, Cosmos, AzureML). His expertise spans deep learning, Bayesian machine learning, computer vision, auction theory, data mining, large scale machine learning, map/reduce computation, and databases.
Highlights
Our Nature Comms paper about AI in the context of early cancer screening is now out 🎉 This is the result of an exciting collaboration with Cyted where we designed AI models that could be used with Cyted's non-invasive screening tool EndoSign®. Our AI models can help to efficiently analyze the histopathology images coming out of EndoSign, and identify goblet cells. We showed that an AI-assisted workflow could significantly reduce pathologists' workloads, paving the way for large-scale screening for patients at risk of developing esophageal cancer.
Paper: https://lnkd.in/eR4vYN52 Blog post summarizing the work: https://msft.it/6041cgyh3
Thanks to the amazing team Kenza Bouzid Harshita Sharma Sarah Killcoyne Daniel Coelho de Castro Maximilian Ilse Valentina Salvatelli, PhD Ozan Oktay Sumanth Murthy Lucas Bordeaux Luiza Moore MD PhD Anja Thieme Hannah Richardson (née Murfet) Aditya Nori Marcel Gehrung, PhD Javier Alvarez Valle
How can we compute better feature representations of medical images, that are accurate and powerful for later use in, for example, large language models? This is the topic of our team's latest preprint "RAD-DINO: Exploring Scalable Medical Image Encoders Beyond Text Supervision." https://lnkd.in/eHZBYGfV
RAD-DINO highlights: 1️⃣ RAD-DINO demonstrates that weak supervision using text data is not essential for learning visual features in multimodal biomedical applications. RAD-DINO employs self-supervision with imaging data only, achieving comparable or even better performance. This means that the vast image-only datasources can be used for further extension of this model. 2️⃣ Superior Correlation with Clinical Data: RAD-DINO's features show a stronger correlation with clinical information, such as patient medical records, without being trained on such. 3️⃣ Scalable Performance: Extensive ablation studies demonstrate that RAD-DINO's performance scales with increased training dataset size, diversity, and higher input resolution.
Great teamwork with Fernando Pérez-García Harshita Sharma Sam Bond-Taylor Kenza Bouzid Valentina Salvatelli, PhD Maximilian Ilse Shruthi Bannur Daniel Coelho de Castro Matthew Lungren MD MPH Maria Teodora Wetscherek Noel C. F. Codella, Ph.D. Stephanie Hyland Javier Alvarez Valle Ozan Oktay
