Vision2023Shipped

Leukemia classification

98.9% held-out

EfficientNetB7 on blood-smear imagery — foundation for later triage work.

Problem

Distinguish Acute Lymphoblastic Leukemia blood smears from healthy controls with enough accuracy to support later screening work where pathologists are scarce.

Constraints

  • Limited labeled imagery — augmentations carried much of the gain
  • False negatives are costly; prototype framed as triage, not diagnosis
  • Deployment realities (device, clearance) deferred to the capstone follow-on

What shipped

An EfficientNetB7 classifier with rotation, flip, and color-jitter augmentations. Accuracy lifted roughly thirteen points; held-out test set reached 98.9%. Became the technical base for the Berkeley MIDS “A World for Every Child” screening prototype.

Outcome

A measurable vision baseline and a clear handoff into a deployment-minded capstone — not a clinical product by itself.

More cases

All works →