Morphome
Mapping the Latent Space of Patient Geometry
Morphome (formerly Cxr8 Explorer) is an educational application that allows users to view individual chest x-ray images, as well as to browse and filter the images based on various criteria such as image quality, patient age, and the presence of specific abnormalities. It includes a model that allows navigation of the latent space of patient geometries, using our “anat-o-mixer” control.
The CXR8 dataset is a collection of chest x-ray images that has been produced by the NIH Clinical Center. It includes a total of 220,000 grayscale images, with a resolution of 1024x1024 pixels. The images have been labeled as either normal or abnormal, with abnormalities ranging from common respiratory diseases such as pneumonia and tuberculosis, to other conditions such as cancer and cardiac abnormalities.
Theory of Operation of the Model
The model of interest arises from an analogy between a locally-connected variational auto-encoder and the mammalian visual system.
- Van Essen’s Distributed Hierarchy proposed that vision happens by hierarchically refiltering with the same basic modules. So if there are intrinsic mechanisms for developing Gabor filter banks, then we should look at stacks of Gabor filter banks to build approximations of mammalian vision.
- Olshausen’s Shifter dynamic routing circuit proposed a bayesian framework for incorporating variations in scene recognition.
- Mumford’s dynamic blackboard model of the thalamus proposed temporal coding as a primary means of altering the condition variables during vision.
Applying these principles to modern ML models (Resnet-50 in particular) has produced interesting results, utilizing much fewer parameters than SOTA models.