Our Research
Fairness & Trustworthiness in Medical AI
Although AI has tremendous potential for automating expert-level diagnoses in radiology, they are susceptible to bias and unfairness against underrepresented groups, which could perpetuate pre-existing healthcare disparities. These and other pitfalls, such as shortcut learning (making the right diagnosis for the wrong reasons), threaten the safe and ethical use of AI in clinical care. We are investigating these biases and pitfalls across different applications in medical imaging and developing techniques to measure and mitigate them.
Recent publications
Vision Transformers
Vision transformers (ViTs) have been proposed as an improvement over convolutional neural networks (CNNs) for general computer vision tasks, but their advantage for radiology image diagnosis is unclear. We performed the first systematic evaluation of ViTs vs. CNNs and found improvements in shortcut learning and disease localization while maintaining similar performance. We are actively investigating further the potential benefits of ViTs in medical imaging AI.
Recent publications
Collaborative Meta Learning (CoMet)
AI models in medical imaging typically focus on a single task or problem. This is a major bottleneck, especially when considering the heterogeneity of datasets and accompanying annotations. We are developing CoMet (Collaborative Meta Learning) — a set of techniques to train single models capable of performing multiple tasks using heterogeneous datasets, even across different modalities and disease labels.
Recent publications
Image Segmentation
Image segmentation is a foundational tool for medical image analysis. We have worked on building segmentation tools for a variety of modalities, body parts, and diseases.
Recent publications
AI-assisted Clinical Decision Support
We are actively building AI solutions for clinical decision support across multiple diseases and modalities.
Recent publications
Virtual Tumor Biopsy
The gold-standard for tumor diagnosis and grading is through invasive biopsy — AI has the potential to use imaging to perform these diagnoses ‘virtually’ without the need for an invasive procedure.
Recent publications