The mitosis count, one of the main components considered for grading breast cancer on histology images, is used to assess tumour proliferation. On breast histology sections, different coloured stains are used to highlight existing cellular components. To keep the wealth of information in the stain colour representation and decrease the sensitivity of detection and classification models to stain variations, the RGB Histogram Specification method is used as a preprocessing step in the training of a modified deep convolutional neural network. Different models are trained on raw and stain normalised images using different databases. Evaluation results show more stable detection performance for various imaging conditions. Combining different data sources and employing stain colour normalisation and transfer learning, a network is trained that can be used for the general mitosis detection task and dealing with staining and scanner variations.
CITATION STYLE
Hamidinekoo, A., & Zwiggelaar, R. (2017). Stain colour normalisation to improve mitosis detection on breast histology images. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 10553 LNCS, pp. 213–221). Springer Verlag. https://doi.org/10.1007/978-3-319-67558-9_25
Mendeley helps you to discover research relevant for your work.