Adaptive-AutoMO: A domain adaptive automated multiobjective neural network for reliable lesion malignancy prediction via digital breast tomosynthesis

Abstract

Early diagnosis of breast cancer remains a significant global health challenge, and the potential use of deep learning in Digital Breast Tomosynthesis (DBT) based breast cancer diagnosis is a promising avenue. To address data scarcity and domain shift problems in building a lesion malignancy predictive model, we proposed a domain adaptive automated multiobjective neural network (Adaptive-AutoMO) for reliable lesion malignancy prediction via DBT. Adaptive-AutoMO addresses three key challenges simultaneously, they are: privacy preserving, credibility measurement, and balance, which consists of training, adaptation and testing stages. In the training stage, we developed a multiobjective immune neural architecture search algorithm (MINAS) to generate a Pareto-optimal model set with balanced sensitivity and specificity and introduced a Bayesian optimization algorithm to optimize the hyperparameters. In the adaptation stage, a semi-supervised domain adaptive feature network based on maximum mean discrepancy (MMD-SSDAF) was designed, which can make the balanced models adaptable to the target domain and preserve the data privacy in the source domain. In the testing stage, we proposed an evidence reasoning method based on entropy (ERE) that can fuse multiple adapted models and estimate uncertainty to improve the model credibility. The experiments on two DBT image datasets (source and target domain datasets) revealed that Adaptive-AutoMO outperformed ResNet-18, DenseNet-121, and other available domain adaptive models. Meanwhile, the removal of high uncertainty samples resulted in a performance improvement in the target domain. These experiments affirmed that Adaptive-AutoMO can not only enhance model’s performance, but also preserve privacy in the source domain data, boost model credibility, and achieve a balance between sensitivity and specificity.