College of Intelligence and Computing, Tianjin University, Tianjin, China; Tianjin Key Laboratory of Cognitive Computing and Application, Tianjin, China; Tianjin Key Laboratory of Advanced Networking, Tianjin, China.
College of Intelligence and Computing, Tianjin University, Tianjin, China; Tianjin Key Laboratory of Cognitive Computing and Application, Tianjin, China; Tianjin Key Laboratory of Advanced Networking, Tianjin, China.
Comput Biol Med. 2022 May;144:105347. doi: 10.1016/j.compbiomed.2022.105347. Epub 2022 Mar 2.
[S U M M A R Y] Weakly supervised segmentation for medical images ease the reliance of models on pixel-level annotation while advancing the field of computer-aided diagnosis. However, the differences in nodule size in thyroid ultrasound images and the limitations of class activation maps in weakly supervised segmentation methods typically lead to under- and/or over-segmentation problems in real predictions. To alleviate this problem, we propose a weakly supervised segmentation neural network approach. This new method is based on a dual branch soft erase module that expands the foreground response region while constraining the erroneous expansion of the foreground region by the enhancement of background features. The sensitivity of this neural network to the nodule scale size is further enhanced by the scale feature adaptation module, which in turn generates integral and high-quality segmentation masks. In addition, while the nodule area can be significantly expanded through soft erase module and scale feature adaptation module, the activation effect in the nodule edge area is still not satisfactory, so that we further add an edge-based attention mechanism to strengthen the nodule edge segmentation effect. The results of experiments performed on the thyroid ultrasound image dataset showed that our new approach significantly outperformed existing weakly supervised semantic segmentation methods, e.g., 5.9% and 6.3% more accurate than the second-based results in terms of Jaccard and Dice coefficients, respectively.
[摘要] 医学图像的弱监督分割减轻了模型对像素级注释的依赖，同时推进了计算机辅助诊断领域的发展。然而，甲状腺超声图像中结节大小的差异和弱监督分割方法中类激活图的局限性通常导致实际预测中的欠分割和/或过分割问题。为了解决这个问题，我们提出了一种弱监督分割神经网络方法。这种新方法基于双分支软擦除模块，通过增强背景特征来扩展前景响应区域，同时限制前景区域的错误扩展。通过尺度特征自适应模块进一步增强了神经网络对结节尺度大小的敏感性，从而生成完整的高质量分割掩模。此外，虽然通过软擦除模块和尺度特征自适应模块可以显著扩展结节区域，但在结节边缘区域的激活效果仍不理想，因此我们进一步添加基于边缘的注意力机制来增强结节边缘的分割效果。在甲状腺超声图像数据集上的实验结果表明，我们的新方法明显优于现有的弱监督语义分割方法，例如在 Jaccard 和 Dice 系数方面，分别比基于第二的结果准确 5.9%和 6.3%。
