Problem
Predix Space works on neovascular age-related macular degeneration (AMD), a common disease whose forms can cause severe, almost irreversible vision loss. The standard treatment is anti-VEGF therapy. It does not always work, can cause complications such as rupture of the retinal pigment epithelium, and is expensive, so a failed course hurts both the patient and the clinic. The team wants to predict in advance whether therapy will work for a given patient and to measure retinal structures quantitatively.
Example from the article: in patients with pigment epithelium detachment (PED), the team classifies the outcome after anti-VEGF as reattached, not reattached or ruptured.
Solution in brief
Annotate retinal structures on optical coherence tomography (OCT) scans and compute their length, height and area. Feed these features to a random forest that predicts the therapy outcome. In parallel, train a U-Net to find the same structures automatically, so the whole chain runs from a raw scan.
Questions this case answers
- How to predict a binary treatment outcome for each object from a small table of interpretable measurements extracted from images by a classical model
- How to replace manual annotation of structures in images with an automatic segmentation model whose only job is to produce the same measurements
- How to predict whether anti-VEGF therapy will work for a patient with neovascular AMD from OCT images
- How to turn manually annotated medical images into quantitative features for a classical ML model
Tools and methods
- LabelMe — manual segmentation of retinal structures (4 biomarkers so far)
- Contour detection and morphometry — length, height and area of each structure
- Random forest — predicts reattachment vs no reattachment of PED
- U-Net — segments biomarkers on OCT scans
- Feature importance — checks which biomarkers drive the prediction
Google or a one-time email code. No password.