Classification vs Deep Learning in Cancer Degree on Limited Histopathology Datasets

Abstract

Today deep learning techniques (DL) are the main focus in classification of disease conditions from histology slides, but this task used to be done by more traditional machine learning pipeline algorithms (MLp). The first can learn autonomously, without any feature engineering. But some questions arise: can we design a fully automated MLp? Can that MLp match DL, at least in some tasks? how should it be designed? Can both be useful and/or complement each other? In this chapter we try to answer those questions. In the process, we design an automated MLp, build DL architectures, apply both to cancer grading, compare accuracy experimentally and discuss the remaining issues. Surprisingly, a carefully designed MLp procedure (acc. 86.5%) compared favorably to deep learning (best acc. 82%) and to humans (acc. 84%) when detecting degree of atypia for breast cancer prognosis on limited-sized publicly available Mytos dataset, with the same DL architectures that achieved accuracies of 97% on a different cancer classification task. Most importantly, we discuss advantages and limitations of alternatives, in particular what features make DL superior and may justify that choice, but also how MLp can be almost fully automated and produce useful structures characterization. Finally, we raise challenges, identifying how MLp and DL should evolve to offer explainability and integrate humans in the loop.