A Pilot Study: Deep Multi-Instance Learning for Origin Tracing of Brain Metastases

Abstract Treatment decisions for brain metastasis heavily rely on identifying the primary site, which is typically accomplished through biomarker-based techniques such as genomics and histopathology. However, limited healthcare resources sometimes can hinder their availability. Therefore, we innovatively transform origin tracing into an image classification task. Based on T1ce-MRI, we develop a non-invasive and cost-effective pipeline, called deep multi-instance learning (DMIL). The DMIL-based pipeline includes three steps: pre-processing, training and testing. Particularly, in pre-processing, mix-modal data decoration is proposed to learn multiple modal knowledge. For DMIL training, center-point-based lesion identification is employed to automatically crop ROIs, eliminating the need for manual intervention. Additionally, self-adaptive lesion classification aims to achieve slice-wise origin tracing. During the inference stage, to address the uncertainty stemming from heterogeneity within a patient's volume, we design a voting majority mechanism to make final patient-wise predictions. Evaluated on the clinical dataset, our DMIL-based pipeline demonstrated promising results. The best patient-wise results achieved at 87.27% (accuracy), 85.00% (PPV) and 83.33% (sensitivity).