Multi-level learning approach for prostate histopathology images classification

The classification of prostate histopathology images is concerned with the identification of multiple classes encompassing different grades of malignancy corresponding to the different observable textural patterns. To address the prostate multiclass classification problem, decomposition schemes are well-known techniques to solve the multiclass classification tasks. Among them two common binary approaches using one-versus-all (OVA) and one-versus-one (OVO) have gained a significant attention from the research community. However, in the case of OVO, the correlation between different classes is not considered as the multiclass problem is broken into multiple independent binary problems. On the other hand, OVA introduces an artificial class imbalance, which degrades the classification performance. In this paper, a new multiclass approach, named multi-level learning architecture (MLA), which handles the binary classification tasks in the multi-level strategy. It does so by taking the correlation between different classes and the domain knowledge into account. In addition, the proposed approach relies upon the ‘divide and conquer’ principle, and work by dividing each binary task into two separate tasks, strong and weak, based on the power of the samples in each binary task. In turn, this motivates the strong samples to produce the final prediction since they have more information about the considered task. Experiments on prostate histopathology images show that the MLA significantly outperforms existing OVO and OVA approaches when they applied on the ensemble framework. The results indicate the high effectiveness of the ensemble framework with MLA scheme in dealing with the prostate multiclass classification problem.