Iterative Annotation to Ease Neural Network Training: Specialized Machine Learning in Pathology

Background:

Automated diagnosis in medical imaging has been hindered by a lack of well-annotated images in large data sets, mainly due to the limited availability and high cost of clinical experts' time. Neural network detection of objects in these images is a potential solution to this problem, but current technologies still require numerous images annotated by human experts before they perform to a reasonable degree of accuracy.  This Catch-22 situation makes the cost of creating computer vision diagnosis software prohibitively expensive.

Technology Overview:

To address this problem UB researchers designed a user-friendly annotation interface which allows intuitive, greatly accelerated neural network training.  Clinicians train the software by annotating a small number of images.  The software then uses this data to predict and overlay annotations on a few new images, which the user can easily correct via 'clicking and dragging' color-coded object boundaries.  After only a few rounds of iterative training, a highly accurate trained AI can be produced and deployed.  The software has broad applications and has already been applied to kidney histology slides and MRI images of human prostate.

Advantages:

​​​​​​​

• Semi-automated image annotation and segmentation
• Intuitive user interface
• Rapid AI development and deployment

Applications:

• Histology/Pathology
• Medical Imaging
• Microscopy

Intellectual Property Summary:

Copyright & Know-how

Stage of Development:

Technology validated in lab, TRL 4

Licensing Status:

Available for license

Publication links:

Nat Mach Intell., 2019 February