Abstract
The early diagnosis of cancer based on histopathology images plays an important role in medical science. Existing techniques generally partition the original histopathology image into small pieces for further classification. However, due to the fact that the number of benign (majority) samples is much larger than that of malignant (minority) samples, the classification is significantly imbalanced which adversely affects classification performance. Undersampling is commonly used to address the class-imbalance problem. However, existing methods are typically time consuming so they are not suitable to handle large-scale and high-dimensional data. In this paper we propose a fast and scalable undersampling method, hashing-based undersampling (HBU), to address class imbalance in large-scale medical image classification. Benign images are hashed and then placed into different buckets according to their locations in the input space. Undersampling is achieved by proportionally selecting benign images from the hash buckets. The HBU method is experimentally evaluated on two real histopathology image datasets, CAMELYON16 and ACDC@LUNGHP, by comparison with existing methods. Experimental results show that the HBU method outperforms six state-of-the-art methods and is scalable and fast.