2021/02/21 by Sara Sangalli, Ertunç Erdil, Sangalli, Sara +7
Computer Science · Health Professions · Medicine · #Artificial Intelligence (cs.AI) #Artificial Intelligence in Healthcare #COVID-19 diagnosis using AI #FOS: Computer and information sciences #Imbalanced Data Classification Techniques #Machine Learning (cs.LG)
paper · pdf · doi:10.48550/arxiv.2102.12894
openalex publication_date 2021/02/21 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Deep neural networks (DNNs) are notorious for making more mistakes for the\nclasses that have substantially fewer samples than the others during training.\nSuch class imbalance is ubiquitous in clinical applications and very crucial to\nhandle because the classes with fewer samples most often correspond to critical\ncases (e.g., cancer) where misclassifications can have severe consequences. Not\nto miss such cases, binary classifiers need to be operated at high True\nPositive Rates (TPRs) by setting a higher threshold, but this comes at the cost\nof very high False Positive Rates (FPRs) for problems with class imbalance.\nExisting methods for learning under class imbalance most often do not take this\ninto account. We argue that prediction accuracy should be improved by\nemphasizing reducing FPRs at high TPRs for problems where misclassification of\nthe positive, i.e. critical, class samples are associated with higher cost. To\nthis end, we pose the training of a DNN for binary classification as a\nconstrained optimization problem and introduce a novel constraint that can be\nused with existing loss functions to enforce maximal area under the ROC curve\n(AUC) through prioritizing FPR reduction at high TPR. We solve the resulting\nconstrained optimization problem using an Augmented Lagrangian method (ALM).\nGoing beyond binary, we also propose two possible extensions of the proposed\nconstraint for multi-class classification problems. We present experimental\nresults for image-based binary and multi-class classification applications\nusing an in-house medical imaging dataset, CIFAR10, and CIFAR100. Our results\ndemonstrate that the proposed method improves the baselines in majority of the\ncases by attaining higher accuracy on critical classes while reducing the\nmisclassification rate for the non-critical class samples.\n