Automated method design for cancer image classification by Differential Evolution and Ensembling

Natalia Oviedo Acosta, Stefan Klein, Martijn P. A. Starmans

Keywords: AutoML, evolutionary algorithms, hyperparameter optimization, computer-aided diagnosis

TL;DR: An AutoML framework using Differential Evolution and ensembling to automate and accelerate deep learning model design for cancer image classification.

Abstract: Developing deep learning models for cancer image classification requires many method design choices, such as in data preprocessing, model architecture, hyperparameters and training procedures. Typically, these are manually tuned, a process that is time-consuming, expert-dependent, and often irreproducible. To address these challenges, we propose an automated machine learning (AutoML) framework that optimizes model design without human intervention. To ensure a comprehensive exploration of diverse architectures and hyperparameter configurations, we define a search space based on state-of-the-art literature in cancer imaging. Our framework employs Differential Evolution and Hyperband (DEHB), which integrates evolutionary search algorithms to balance search space exploration and exploitation, combined with adaptive resource allocation to mitigate the high computational cost of training multiple models. To enhance model robustness and reduce overfitting in data-limited scenarios, we incorporate ensembling. We validate our approach on four public cancer classification datasets encompassing 750 patients with either MRI or CT. The proposed framework demonstrates higher performance when compared to a DenseNet-121 baseline. While exploring multiple configurations, our approach reduces training time by a factor of two to five compared to the baseline. By automating model design and improving generalization across datasets, our framework has substantial potential for broad applications across cancer imaging, thereby streamlining deep learning model development.

Submission Number: 4