Iterative Ensemble Transductive Learning for Microscopy Image Analysis

In automatic histopathology and microscopy image analysis, due to high patient-level variability, the model trained based on the images from a set of patients may not perform well on the images from another set of patients. To overcome this issue, motivated by transductive learning and ensemble learning, we propose an iterative framework to train ensemble transductive models using pseudo-labels of test data. In each iteration, a number of individual models are first trained by combining the training set with part of randomly selected test data which have high prediction confidence, and then ensembled to predict the labels of test set for the next iteration. In this way, the latent information in test set would be exposed to the model such that the model can directly learn from the test data. Experimental evaluation on the white blood cancer microscopic image set and the breast histopathology image set shows that the proposed approach significantly outperforms the traditional ensemble models.