Multiple Instance Learning for Lymph Node Metastasis Prediction from Cervical Cancer MRI

Abstract

Lymph node metastasis (LNM) is an important prognostic factor for recurrence and overall survival of cancer patients. The current LNM diagnosis is based on histopathologic examination after surgical lymphadenectomy, but an accurate and noninvasive method for LNM diagnosis is essential in selecting reasonable surgical operations and treatment plans. This paper presents an attention based multiple instance learning (MIL) model to diagnose LNM from cervical cancer multimodal MRI. The proposed MIL model adopts convolutional neural network (CNN) to extract features from multimodal MRI and attention-based pooling to make patient-level LNM status prediction. By incorporating the MIL and attention mechanism, the top rank MRI slice with informative regions in each LNM positive patient is visualized to provide the interpretability for LNM diagnosis. Experiments evaluated on a cohort of 241 cervical cancer patients show improvements in LNM status prediction compared with existing comparative models, which indicates the advantages of our designed model.