Pathological Image Classification Based on Hard Example Guided CNN

The diagnosis of biopsy tissue with hematoxylin and eosin (H&E) stained images has been widely used by pathologists to detect the lesions and assess the malignancy. Nevertheless, the diagnostic result relies on the visual observation of pathologists which may vary from person to person under different circumstances. With the advantage of automatically and adaptively learning features at multiple levels of abstraction, Convolutional Neural Networks (CNNs) have rapidly become promising alternatives for pathological image analysis. Therefore, in this paper, we propose an effective method for tumor classification called Hard Example Guided CNN. Our contribution is twofold: firstly, to optimize image representation, we design the CNN architecture as dual-branch, used for extracting global features and local features simultaneously. Secondly, we propose a re-weight training algorithm, which improves learning accuracy and accelerates the convergence by increasing the weight of hard examples. Extensive experiments on multiple datasets demonstrate the superiority of our proposed classification method.