SOCR logoBPAD

Medical Image Augmentation Explorer

Apply rotation, translation, scale, noise, intensity shift and horizontal flip to a synthetic CT slice and see which transforms destroy the label.

  • augmentation
  • data pipeline
  • Hounsfield units
  • laterality
Zero-padded fraction
0.031
border pixels with no source data
Mean HU
36.595
SD HU
19.409

Augmented image

HU (black=low, white=high)

Intensity histogram

waterfat/lesion range-150-100-5005010015005001000150020002500Hounsfield UnitsCount
Would a radiologist still trust this image?
Rotation and scaling introduce zero-padded regions that carry no anatomical meaning — a network can learn to key on them. A horizontal flip silently reassigns laterality. Intensity shifts of a hundred HU push soft tissue toward fat or bone ranges. Augmentation only helps when it stays inside the space of physically plausible acquisitions.