Advanced Medical Imaging Research
We are pioneering physics-grounded, label-free approaches for diagnostic AI.
The Annotation Bottleneck
Medical imaging annotation remains a critical bottleneck in radiology AI development, requiring extensive expert time and limiting scalability. Current supervised-learning pipelines depend on large volumes of labeled data — annotations produced by subspecialist radiologists whose time is scarce and whose inter-observer variability sets an inherent ceiling on model accuracy.
Our Active Research
Three converging research directions toward physics-grounded, label-free diagnostic AI.
Label-Free Anomaly Detection
Our Imaging Anomaly Mapper learns the statistical distribution of healthy tissue patterns using normative modeling with PCA and Autoencoder architectures — identifying deviations without requiring annotated datasets.
K-Space Physics Modeling
We are developing approaches that leverage k-space (frequency domain) MRI data to extract molecular-level diagnostic signatures invisible to standard spatial-domain methods.
DICOM Voxel Representations
By operating directly on DICOM voxel intensities as physical measurements rather than visual features, we aim to enable non-invasive detection of biochemical tissue changes at unprecedented resolution.
Imaging Anomaly Mapper
Our Imaging Anomaly Mapper addresses the annotation bottleneck through label-free anomaly detection using normative modeling with PCA and Autoencoder architectures. By learning the statistical distribution of healthy tissue patterns, the system identifies deviations without requiring annotated datasets — enabling rapid diagnostic model development.