Object Knowledge-Aware Multiple Instance Learning for Small Tumor Segmentation

Abstract

In image-guided radiofrequency ablation therapy, the accurate segmentation of small tumors is critical for ensuring effective treatment and improving patient prognosis. However, existing segmentation methods cannot characterize the geometric shape of tumors and the differences between tumors and surrounding tissues well due to the scarcity of precise annotations, various sizes, and boundary ambiguity. To address these challenges, we propose the object knowledge-aware multiple instance learning (OKMIL) method, which integrates clinical tumor knowledge into the MIL framework. First, an instance classification network is designed to generate a precise tumor mask and a tight tumor box. This classification network incorporates two innovative components: 1) a tumor background-aware module that is designed to address the noise labels within the loose tumor box by learning the feature difference between the tumors and surrounding tissues; and 2) a tumor scale-aware module that is designed to alleviate the missed detection of small tumors by capturing the long-range dependencies at the feature level. Second, a shape-aware MIL loss is designed to address the large shape variations by uncertainty estimation of tumor instances during training. Validation experiments are conducted on the liver tumor segmentation (LiTS) dataset and nasopharyngeal neoplasm segmentation (NpNS) dataset. The evaluation results demonstrate that our method can achieve higher accuracy and fewer missed detections than existing weakly supervised methods with box-level annotations. In particular, in the segmentation of tiny and small liver tumors, the dice coefficient improves by 14% and 13%, respectively.

Note:
Funding declaration: This research is financially supported by National Natural Science Foundation of China (No.62372358, No.82272735, No.62302355), is support by the Key Research and Development Plan of Shaanxi Province (No.2024GH-ZDXM-35), is supported by the Natural Science Basic Research Program of Shaanxi Province (No.2023-JC-QN0719), is supported by the Basic and Applied Basic Research Foundation of Guangdong Province (No.2022A1515110453), is supported by Technology Project of Xianyang City (Key Research and Development Plan) (No.JBGS-013), and is supported by Advanced Interdisciplinary Studies Project of the Second Affiliated Hospital of Air force medical university (No.2021TYJC-004).

Conflict of Interests: The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Keywords: Multiple instance learning, Tumor clinical knowledge, Shape-aware learning, Background-aware learning, Small tumor segmentation