Outlier Robust Disease Classification via Stochastic Confidence Network

Abstract

Accurate and timely diagnosis and classification of diseases using medical imaging data are essential for effective treatment planning and prognosis. Yet, the presence of outliers, which are rare and distinctive data samples, can result in substantial deviations from the typical distribution of a dataset, particularly due to atypical or uncommon medical conditions. Consequently, outliers can significantly impact the accuracy of deep learning (DL) models used in medical imaging-based diagnosis. To counter this, in this work, we propose a novel DL model, dubbed the Stochastic Confidence Network (SCN), designed to be robust to outliers. SCN leverages image patches and generates a decoded latent matrix representing high-level categorical features. By performing a stochastic comparison of the decoded latent matrix between outliers and typical samples, SCN eliminates irrelevant patches of outliers and resamples outliers into a typical distribution, thereby ensuring statistically confident predictions. We evaluated the performance of SCN on two databases for diagnosing breast tumors with 780 ultrasound images and Alzheimer’s disease with 2,700 3D PET volumes, with outliers present in both databases. Our experimental results demonstrated the robustness of SCN in classifying outliers, thereby yielding improved diagnostic performance, compared with state-of-the-art models, by a large margin. Our findings suggest that SCN can provide precise and outlier-resistant diagnostic performance in breast cancer and Alzheimer’s disease and is scalable to other medical imaging modalities. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2023.