training on patches

A commonly adopted approch is training on image patches that are equally sampled from each class[1]. This, however, biases the classifier towards rare classes and may result in over-segmentation. To counter this, [2] proposes to train a second CNN on samples with a class distribution close to the real one, but oversample pixels that were incorrectly classified in the first stage. In practice, two stage training schemes can be prone to overfit- ting and sensitive to the state of the first classifier.

[1] K. Kamnitsas et al., “Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation,” Medical Image Analysis, vol. 36, pp. 61–78, Feb. 2017. [2]D. C. Cireşan, A. Giusti, L. M. Gambardella, and J. Schmidhuber, “Mitosis Detection in Breast Cancer Histology Images with Deep Neural Networks,” in Advanced Information Systems Engineering, vol. 7908, C. Salinesi, M. C. Norrie, and Ó. Pastor, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2013, pp. 411–418.