AI-Based Small Blob Detector for Imaging Biomarker Quantification

Imaging biomarkers are being increasingly applied for early diagnosis and staging of disease in humans. Detecting and segmenting objects from images are often the first steps in imaging biomarker quantification. Of particular interest in this research is glomeruli segmentation from kidney MR images - a small object (blob) detection problem. It has unique challenges: (1) the size of glomerular is extremely small and similar to noises from images; (2) there are massive of glomeruli in kidney (over 1 million glomeruli in human kidney), and the intensity distribution is heterogenous; (3) a large portion of glomeruli are overlapping and touched in images.

To address these challenges, we have developed a suite of deep learning based small blob detectors. The first small blob detector is jointly constrained deep learning and Hessian analysis. The theoretical foundation of Hessian analysis guarantees that pre-segmentation will recognize all true convex blobs and non-blob convex objects, resulting in a blob superset. Joining effort from deep learning model (U-Net) allows us to distinguish true blobs from the superset and alleviate the over-detection issue from Hessian analysis. As an extension of this detector, we develop bounded scales for blobs to be transformed to local optimum Difference of Gaussian (DoG) space adaptively to improve the computational efficiency and segmentation accuracy. The last part of the talk is our recent work on GAN models for imaging synthesis to address limited data challenge often seen in medical applications. The experimental results from the blob detectors show the potential to automatically detect glomeruli, enabling new measurements of renal microstructures in preclinical and clinical studies.