To accurately segment various clinical lesions from computed tomography (Computed Tomography, CT) images is a critical task for the oncology imaging. However, researchers design existing segmentation frameworks for lesions of a specific disease, as well as segmenting visually inconspicuous and small-scale tumors remains a challenging problem. Thus , this paper imitates the diagnostic behavior of clinicians and proposes a non-significant small tumor segmentation framework based on multi-scale visual information and non-local target mining. This framework first combines scale space theory to extract differentiated features at 1.0X, 0.5X, and 1.5X scales , and then applies scale fusion module to hierarchically fuse feature maps of specific scales for obtaining a comprehensive and accurate tumor characterization. The obtained features capture the long-range semantic dependencies of channels and spatial locations through global localization module, locate tumors from a global perspective, obtaining initial prediction results. Layer focusing module performs context exploration based on foreground and background features, focuses on error areas layer by layer, and uses element-by-element addition and subtraction to eliminate these errors. By gradually refining the initial prediction results, framework finally achieves more refined tumor segmentation results. Empirical experiments on the small intestinal stromal tumor dataset (SISD) and the pancreatic tumor dataset (PTD) show that the proposed framework outperforms 10 existing state-of-the-art methods in 6 standard metrics. Our framework achieves 58.37% and 57.64% (Dice) on the SISD and PID respectively, which are 7.38% and 4.07% higher than the previous best results. Author's github homepage will publish the code.