We developed a fully unsupervised framework for image segmentation and investigated its performance on natural images and more importantly, two medical image datasets. The proposed framework combines intra-image contrastive learning and local patch reconstruction to jointly learn a structured latent space. Then, it coarse-grains the latent pixel-level embeddings into a series of coarse-to-fine segmentations at various levels of granularity. The proposed method is end-to-end unsupervised and outperforms existing unsupervised segmentation methods in our quantitative experiments.

We developed and optimized a deep learning algorithm to produce Gadolinium contrast in mouse and human brain MRI directly from a single non-contrast structural MRI. We showed sensible prediction results by quantitatively comparing not only the contrast images but also the downstream scientific findings with the ground truths on two species, multiple studies, and various levels of brain disorders.

We proposed a generalizable adaptation of Focal Loss to keypoint detection leveraging difficulty scores from a discriminator.