Deep learning introduces a family of powerful algorithms that can help to discover features of disease in medical images, and assist with decision support tools. In the context of medical imaging, there are several interesting challenges:
- ~1500 different imaging studies
- Many distinct imaging modalities (e.g., DR, CT, MR, US, NM)
- Visualizing a variety of body regions (e.g., head, chest, extremities, abdomen)
- Each containing dozens of organs and body parts (e.g., liver, pancreas, kidney)
- Each affected by hundreds of diseases (Total: ~23,000 conditions)
- Typically 12- or 16-bit gray scale, rather than color images.
- 3D volumetric data sets, sometimes time as 4th dimension.
- Multi-channel data (e.g., MR T1-weighing, T2-weighting, flow-sensitive, post-contrast).
- Differential diagnosis varies by anatomic structure, requiring anatomic segmentation.
These challenges are further complicated by the sheer size of the data that is available at Stanford Medicine (2016):
The Langlotzlab has several projects that are testing these algorithms to assist with tasks relevant to diagnostic radiology: