A bag-of-features approach for malignancy detection in breast histopathology images

The paper addresses the problem of detecting malignancy in breast histopathology images. The proposed method uses bag-of-features method to represent visual content of the image dataset. The features are extracted from training dataset by using SIFT technique. Further classification is carried out using Euclidean distance measure. The method was evaluated using the dataset of breast histopathology images having annotated as normal, invasive and in-situ. The keypoints are extracted from R, G and B components and collectively form discriminative visual features. In the classification task, the paper proposes two-stage classification. In stage-I sample is classified as normal or cancerous. In stage-II, if the sample is annotated as cancerous, it gets further classified according to its type that is, either invasive or in-situ. The method showed stage-I performance as 74.28% for normal and 90.39% for cancerous images. In Stage-II, accuracy for invasive and in-situ types is 98.86% and 84.09% respectively.