Fig. 1
Machine learning analysis of quantitative tissue composition metrics. Digitized hematoxylin and eosin-stained slides were used to optimize machine learning-based tissue classification scripts. A custom-built, random forest, tissue classifier algorithm (Indica Labs, Albuquerque, NM) was trained by two pathologists to develop an optimized, 85-datapoint, tissue classifier script. By annotating regions on randomly selected representative H&E-images comprised of epithelium, stroma, and adipose tissue (A), the random forest algorithm was trained to identify, segment, and quantify areas (in mm.2) on each slide comprised of epithelium (42-datapoints), stroma (37-datapoints), and adipose tissue (6-datapoints) as shown on (B) (Red: epithelium; Green: stroma; Yellow: adipose tissue). C and D show high-power views of the machine’s capacity to identify regions on the slide comprised of adipose tissue (C) as well as epithelium and stroma (D)