Development and Validation of Explainable Artificial Intelligence System for Automatic Diagnosis of Cervical Lymphadenopathy From Ultrasound Images

Abstract

Clinical diagnosis of cervical lymphadenopathy (CLA) using ultrasound images is a time-consuming and laborious process that heavily relies on expert experience. This study aimed to develop an intelligent computer-aided diagnosis (CAD) system using deep learning models (DLMs) to enhance the efficiency of ultrasound screening and diagnostic accuracy of CLA. We retrospectively collected 4089 ultrasound images of cervical lymph nodes across four categories from two hospitals: normal, benign CLA, primary malignant CLA, and metastatic malignant CLA. We employed transfer learning, data augmentation, and five-fold cross-validation to evaluate the diagnostic performance of DLMs with different architectures. To boost the application potential of DLMs, we investigated the potential impact of various optimizers and machine learning classifiers on their diagnostic performance. Our findings revealed that EfficientNet-B1 with transfer learning and root-mean-square-propagation optimizer achieved state-of-the-art performance, with overall accuracies of 97.0% and 90.8% on the internal and external test sets, respectively. Additionally, human–machine comparison experiments and the implementation of explainable artificial intelligence technology further enhance the reliability and safety of DLMs and help clinicians easily understand the DLM results. Finally, we developed an application that can be implemented in systems running Microsoft Windows. However, additional prospective studies are required to validate the clinical utility of the developed application. All pretrained DLMs, codes, and application are available at https://github.com/YubiaoYue/DeepUS-CLN.