Noise-Tolerant Deep Learning for Histopathological Image Segmentation

Abstract

Developing an effective algorithm based on the handcrafted features from histological images (histo-images) is difficult due to the complexity of histo-images. Deep network models have achieved promising performances, as it is capable of capturing high-level features. However, a major hurdle hindering the application of deep learning in histo-image segmentation is to obtain large ground-truth data for training. Taking the segmentations from simple off-the-shelf algorithms as training data will be a new way to address this hurdle. The output from the off-the-shelf segmentations is considered to be noisy data, which requires a new learning scheme for deep learning segmentation. Existing works on noisy label deep learning are largely for image classification. In this thesis, we study whether and how integrating imperfect or noisy “ground-truth” from off-the-shelf segmentation algorithms may help achieve better performance so that the deep learning can be applied to histo-image segmentation with the manageable effort.

Two noise-tolerant deep learning architectures are proposed in this thesis. One is based on the Noisy at Random (NAR) Model, and the other is based on the Noisy Not at Random (NNAR) Model. The largest difference between the two is that NNAR based architecture assumes the label noise is dependent on features of the image. Unlike most existing works, we study how to integrate multiple types of noisy data into one specific model. The proposed method has extensive application when segmentations from multiple off-the-shelf algorithms are available. The implementation of the NNAR based architecture demonstrates its effectiveness and superiority over off-the-shelf and other existing deep-learningbased image segmentation algorithms.