Impact of data preprocessing and augmentation on tumor core segmentation using convolutional neural networks

The relevance of detecting and treating breast cancer in the early stages remains high. In 2020, more than 65,000 new cases of breast cancer were registered, with an average annual growth rate being 2%. Every year, the number of recorded cases of breast cancer is leading in statistics of cancer diseases. The goal of the work is to evaluate the impact of preprocessing and augmentation methods on data sets for segmenting tumor nuclei in medical images under conditions of limited data volume. The experiments use one initial data set and eight variants of its pre-processing using image slicing algorithms to train two models of convolutional neural networks, U-net and U-net with the addition of a ResNet50 encoder. Assessing the quality of neural network training and kernel segmentation is performed using two target metrics, Dice and IoU, as well as by comparing the true location of kernel labels and segmented kernels using neural networks. As a result of training the models on pre-processed data sets, values of target metrics are obtained for the two models for each dataset, including the original one. For the U-net architecture, the Dice and IoU values are 0.742 and 0.5921, for the U-net_ResNet50 architecture they are 0.7458 and 0.5971.