A Morphological Image Preprocessing Method Based on the Geometrical Shape of Lesions to Improve the Lesion Recognition Performance of Convolutional Neural Networks
A Morphological Image Preprocessing Method Based on the Geometrical Shape of Lesions to Improve the Lesion Recognition Performance of Convolutional Neural Networks
复制标题
DOI：
10.1109/access.2022.3187507
复制
发表时间：
2022-01-01
期刊：
影响因子：
3.9
通讯作者：
Kimori, Yoshitaka
中科院分区：
文献类型：
--
作者：
Kimori, Yoshitaka
卷积神经网络(CNN)在计算机视觉相关的医学成像任务中扮演着重要的角色。然而，对于CNN模型来说，数据集中的原始图像的质量可能不足以学习目标对象的特征。当输入图像包含复杂背景时，CNN模型将重点放在对病变识别不重要的区域上，例如背景结构，从而导致输出预测不那么准确。本文提出了一种基于数学形态学的图像预处理方法，该方法利用了病变形状的先验知识，可以有效地解决这一问题。该方法由基于病变区域几何形状信息的h-穹顶变换和随后的图像直方图修正过程组成，具有从背景中选择性增强病变区域的能力。这允许创建明确表示要通过CNN模型学习的重要区域的图像。对胸部X光图像中的肺结节分类和皮肤镜图像中的皮肤病变区域分割的实验表明，与在原始数据集上训练的CNN模型相比，该方法在预处理后的数据集上训练的CNN模型取得了显著的性能改善。
Convolutional neural networks (CNNs) play an important role in computer vision-related tasks for medical imaging. However, the quality of raw images in the dataset can be insufficient for the CNN model to learn the features of the target object. When the input image contains a complex background, the CNN model focuses on regions that are not essential for lesion recognition, such as background structures, leading to less accurate output prediction. This paper shows that this problem can be efficiently solved by an image preprocessing method based on mathematical morphology, which uses a priori knowledge about the lesion shape. The proposed method consists of h-dome transformation based on the geometrical shape information of the lesion region, and subsequent image histogram-modification processes, and has the ability to selectively enhance the lesion region from the background. This allows for the creation of images that explicitly represent the important region to be learned by the CNN model. Experiments on pulmonary nodule classification in chest x-ray images and skin lesion region segmentation in dermatoscopic images demonstrate that theCNNmodels trained on the preprocessed dataset created by the proposed method achieve remarkable performance improvements compared to the CNN models trained on the original dataset.