Ensemble Optimization for Invasive Ductal Carcinoma (IDC) Classification Using Differential Cartesian Genetic Programming

Abstract

The high cost of acquiring annotated histological slides for breast specimens entails exploiting an ensemble of models appropriately trained on small datasets. Histological Image Classification ensembles strive to accurately detect abnormal tissues in the breast samples by determining the correlation between the predictions of its weak learners. Nonetheless, the state-of-the-art ensemble methods, such as boosting and bagging, count merely on manipulating the dataset and lack intelligent ensemble decision making. Furthermore, the methods mentioned above are short of the diversity of the weak models of the ensemble. Likewise, other commonly used voting strategies, such as weighted averaging, are limited to how the classifiers’ diversity and accuracy are balanced. Hence, In this paper, we assemble a Neural Network ensemble that integrates the models trained on small datasets by employing biologically-inspired methods. Our procedure is comprised of two stages. First, we train multiple heterogeneous pre-trained models on the benchmark Breast Histopathology Images for Invasive Ductal Carcinoma (IDC) classification dataset. In the second meta-training phase, we utilize the differential Cartesian Genetic Programming (dCGP) to generate a Neural Network that merges the trained models optimally. We compared our empirical outcomes with other state-of-the-art techniques. Our results demonstrate that improvising a Neural Network ensemble using Cartesian Genetic Programming transcended formerly published algorithms on slim datasets.

Keywords

• Cartesian genetic programming
• hyparameters optimization
• ensemble
• CNNs
• histolpathaological image classification