(Tuesday, 7/16/2019) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 6
Purpose: Deep learning algorithm can be used to identify metastatic tissue in histopathologic scans of lymph node sections. Inspired by the recent success of data augmentation algorithm, this work aims to explore the data augmentation approach to improve the performance of deep learning for the classification task.
Methods: Recently, deep learning based image classification provides state-of-the-art performance, with the comparison to handcrafted features-based shallow classifiers. However, a very large-scale labeled dataset is needed to train the deep neural network, which is laborious to obtain. In our experiment, we explore to use of data augmentation, named mixup, with the goal to constructs virtual training examples. In more detail, mixup can train the deep neural network on the convex combinations of pairs of input and their labels. Despite its simplicity, the mixup data augmentation methods have demonstrated state-of-the-art performance on many datasets. Similar to creating inter-class, mixup increases the robustness of deep neural network when the samples contain corrupted labeled ones.
Results: To evaluate the performance of mixup-based data augmentation method, we evaluate our approach on the binary classification task of Camelyon17 challenge. Using the DenseNet as the classifier, we train the model from scratch on the dataset. Using 5 fold cross-validation, we obtain an accuracy of 88.6 and 90.3 for DenseNet without mixup and the DenseNet with mixup respectively.
Conclusion: We explore to use mixup-based data augmentation for the histopathology patch-classification model that outperforms a deep neural network without mixup-based data augmentation. We find that inter-class can improves the reliability of the model, motivating the application and further research of data augmentation in the medical image analysis domain.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
