Shape-Background Joint-Aware Multiple Instance Learning for Small Tumor Segmentation

作     者：Liu, Haofeng Gou, Shuiping Zhou, Yanyan Jiao, Changzhe Liu, Wenbo Shi, Mei Luo, Zhonghua 

作者机构：Key Laboratory of Intelligent Perception and Image Understanding Education Ministry of China School of Artificial Intelligence Xidian University Xi’an710071 China Department of Interventional Radiology Tangdu Hospital Air Force Medical University Xi’an710038 China Central Laboratory of the First Affiliated Hospital Weifang Medical University Weifang261044 China Department of Radiation Oncology Xijing Hospital Air Force Medical University Xi’an710000 China Fourth Military Medical University China 

出 版 物：《SSRN》 

年 卷 期：2024年

核心收录：

主　　题：Supervised learning 

摘      要：The segmentation of small tumors with irregular shapes and ill-defined boundaries is a crucial step in needle biopsy and radiotherapy. However, manual segmentation is costly and error prone. And acquiring large-scale high-quality annotations for automatic tumor segmentation methods based fully supervised learning is therefore difficult. To address these problem, shape-background joint-aware multiple instance learning (SBJMIL) method is proposed in this paper for small tumor segmentation. In SBJMIL, an instance classification network based on tumor background-aware is designed to deal with the precise annotation scarcity, which can generate tight bounding box annotations via remove background pixels inside loose bounding boxes. Furthermore, a tumor shape-aware module is explored to deal with the class imbalance related to small tumors with irregular shapes, which can learn the morphology information of the tumor via estimating the proportion of positive instances in the prediction results. Quantitative and validation experiments are conducted on the liver tumor segmentation (LiTS-ISBI2017) dataset and nasopharyngeal neoplasm segmentation (NpNS) dataset. For tumor segmentation, the performance of SBJMIL which trained with only loose bounding box annotations is approach to that of fully supervised segmentation methods. And the the segmentation performance of tiny and small liver tumors was improved by 14% and 13% compared to state-of-the-art weakly supervised segmentation methods, respectively. © 2024, The Authors. All rights reserved.