Medical image classification under class imbalance

Thumbnail Image
Zhang, Chuanhai
Major Professor
Carl K. Chang
Committee Member
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Organizational Unit
Journal Issue
Is Version Of
Computer Science

Many medical image classification tasks have a severe class imbalance problem. That is images of target classes of interest, e.g., certain types of diseases, only appear in a very small portion of the entire dataset. These medical image classification tasks share two common issues. First, only a small labeled training set is available due to the expensive manual labeling by highly skilled medical experts. Second, there exists a high imbalance ratio between rare class and common class. The common class occupies a high percentage of the entire dataset and usually has a large sample variety, which makes it difficult to collect a good representative training set for the common class. Convolutional Neural Network (CNN) is currently a state-of-the-art method for image classification. CNN relies on a large training dataset to achieve high classification performance. However, manual labeling is costly and may not even be feasible, which limits CNN from offering high classification performance in practice. This dissertation addresses these two challenging issues with the ultimate goal to improve classification effectiveness and minimize manual labeling effort by the domain experts.

The main contributions of dissertation are summarized as follows. 1) We propose a new real data augmentation method called Unified LF&SM that jointly learns feature representation and a similarity matrix for recommending unlabeled images for the domain experts to verify in order to quickly expand the small labeled training set. Real data augmentation utilizes realistic unlabeled samples rather than synthetic samples. The key of real data augmentation is how to design an effective strategy to select representative samples for certain classes quickly from a large realistic unlabeled dataset. 2) We investigate the effectiveness of six different data augmentation methods and perform a sensitivity study using training sets of different sizes, varieties, and similarities when compared with the test set. 3) We propose a Hierarchical and Unified Data Augmentation (HuDA) method to collect a large representative training dataset for the common class. HuDA incorporates a class hierarchy: class differences on the high level (between the rare class and the common class) and class differences on the low level (between sub-classes of the rare class or the common class). HuDA is capable of significantly reducing time-consuming manual effort while achieving quite similar classification effectiveness as manual selection. 4) We propose a similarity-based active deep learning framework (SAL), which is the first approach to deal with both a significant class imbalance and a small seed training set as far as we know.

Broader Impact: Triplet-based real data augmentation methods utilize the similarity between samples to learn a better feature representation. These methods aim to guarantee that the computed similarity between two samples from the same class is always bigger than the computed similarity between two samples from two different classes. First, our sensitivity study on six different data augmentation methods shows that triplet-based real data augmentation methods always offer the largest improvement on both the recommendation accuracy and the classification performance. These real data augmentation methods are easily extendable to other medical image classification tasks. Our work provides useful insight into how to choose a good training image dataset for medical image classification tasks. Second, to the best of our knowledge, SAL is the first active deep learning framework that deals with a significant class imbalance. Our experiments show that SAL nearly obtains the upper bound classification performance by labeling only 5.6% and 7.5% of all images for the Endoscopy dataset and the Caltech-256 dataset, respectively. This finding confirms that SAL significantly reduces the experts’ manual labeling efforts while achieving near optimal classification performance. SAL works for multi-class image classification and is easily extendable to other medical image classification tasks as well.

Wed May 01 00:00:00 UTC 2019