Enhancing patient-level diagnosis on MR images: a multi-instance learning framework with contributive feature mining

Abstract Interpreting magnetic resonance (MR) images to determine tumor types is a both time-consuming and challenging task for radiologists. Existing computer-aided diagnosis (CAD) models typically focus on enhancing feature extraction from individual slices, despite the fact that a complete MR image sequence consists of a variable number of MR imaging (MRI) slices that could be used together to improve accuracy. To address this issue, we propose a patient-level diagnosis framework, named the Multi-Instance Learning (MIL) method integrated with contributive feature mining (CFM-MIL) which exploits the complete set of MRI slices. The proposed framework treats the MRI case as a bag and the corresponding image patches as instances. It combines two stages of supervised learning phases. In the instance-level feature extraction phase, a Dual-Encoder structure is employed to embed the multi-level features of the instances. During the bag-level feature extraction phase, the MIL model is used to aggregate the instance-level features by modeling the relations of the instances. We elaborate a Siamese structure (Teacher–Student) to guide the MIL model in mining the contributive elements and in diminishing the impact of the less essential elements within the instance-level feature vector. In the testing phase, a similarity-variance-based strategy is proposed to filter out the useless instances within the test dataset. The experimental results on two MR image datasets demonstrate that our proposed CFM-MIL framework can emphasize useful information from MRI data, helping to promote patient-level diagnostic performance.