ML-DSVM+: A meta-learning based deep SVM+ for computer-aided diagnosis

•A novel meta-learning based deep neural network SVM+ (ML-DSVM+) algorithm is proposed for CAD•ML-DSVM+ integrates the bi-channel DNNs and SVM+ into a unified framework for optimization•A new coupled hinge loss is proposed to perform supervised bidirectional transfer•ML-DSVM+ can effectively alleviate the issues of class imbalance and overfitting Transfer learning (TL) can improve the performance of a single-modal medical imaging-based computer-aided diagnosis (CAD) by transferring knowledge from related imaging modalities. Support vector machine plus (SVM+) is a supervised TL classifier specially designed for TL between the paired data in the source and target domains with shared labels. In this work, a novel deep neural network (DNN) based SVM+ (DSVM+) algorithm is proposed for single-modal imaging-based CAD. DSVM+ integrates the bi-channel DNNs and SVM+ classifier into a unified framework to improve the performance of both feature representation and classification. In particular, a new coupled hinge loss function is developed to conduct bidirectional TL between the source and target domains, which further promotes knowledge transfer together with the feature representation under the guidance of shared labels. To alleviate the overfitting caused by the increased parameters in DNNs for limited training samples, the meta-learning based DSVM+ (ML-DSVM+) is further developed, which designs randomly selecting samples from the training data instead of other CAD tasks for meta-tasks. This sampling strategy also can avoid the issue of class imbalance. ML-DSVM+ is evaluated on three medical imaging datasets. It achieves the best results of 88.26±1.40%, 90.45±5.00%, and 87.63±5.56% on accuracy, sensitivity and specificity, respectively, on the Bimodal Breast Ultrasound Image dataset, 90.00±1.05%, 72.55±3.87%, and 96.40±2.26% of the corresponding indices on the Alzheimer's Disease Neuroimaging Initiative dataset, and 85.76±3.12% of classification accuracy, 88.73±7.22% of sensitivity, and 82.60±1.56% of specificity for the Autism Brain Imaging Data Exchange dataset.