Fig. 3.

DTALE provides falsifiable, meaningful and quantitative explanations of nucleus detection model decisions. Unlike other approaches, DTALE can provide explanations that reference object-level morphological measurements such as nuclear size, shape, staining intensity, chromatin texture, perinuclear cytoplasmic staining, etc. In fact, DTALE can use any set of measurable features that make sense to a pathologist to provide quantitative decision tree approximations for black-box classification model decisions. These explanations include global decision criteria, e.g. ‘tumor nuclei are large and have irregular shapes’, as well as decision criteria for individual nuclei of interest