In 2017

The liver has a complex microscopic structure. Alterations at this level affect the macroscopic properties of the liver and overall health of the human body.

Before the non-invasive techniques were validated, physicians routinely took a small liver sample and examine it under microscope. This examination is still subjective.  The goal of this small project was to explore ways in which the stained liver sample can be measured quantitatively, in a fairly precise manner. 

Stained liver sample at 40x magnification. The large transparent bubbles on the right are the targeted features.

For this project I had few labeled samples but enough to apply some Deep Learning magic. Solution was a U-net variant but with reduced capacity. Core elements (scale down, choke, scale up, inter branch shortcuts) were kept. 

I also tried to devise a classical ML segmentation pipeline where candidate objects were filtered by a classifier. Despite putting some non trivial effort in this pipeline, the DL approach was a clear winner in terms of engineering effort and performance. But only when enough labeled data was available. Maybe for really resource tight applications [data+compute], designing a classical CV/ML pipeline makes sense.

U-net architecture. Nothing fancy except that is shallower and with less channels than classical recipe.

For more details on performance and challenges, go to “Deep convolutional neural nets for objective steatosis detection from liver samples” paper.