The presence of artifacts limits the accuracy of detecting skin lesions. The current study presents an extensive appraisal of the impact of eight existing image-segmentation methods on the performance of 10 deep-learning-based models to detect and classify lesions. An empirical review was conducted using dual experimentation- with unsegmented original images, and with segmented images processed using eight segmentation methods on four skin lesion datasets. The learner’s performance was assessed using standard evaluation measures. The results show superior classification performance, achieved with segmented images compared to original images. Otsu’s Binarization approach with ResNet50 model outperforms with an accuracy of 91.9% on ISIC2017 dataset.

伪影的存在限制了检测皮肤损伤的准确性。目前的研究广泛评估了八种现有图像分割方法对 10 种基于深度学习的病变检测和分类模型性能的影响。使用双重实验进行了实证审查——使用未分割的原始图像，以及使用八种分割方法处理的分割图像，对四个皮肤损伤数据集进行了处理。使用标准评估措施评估学习者的表现。结果表明，与原始图像相比，分割图像具有更好的分类性能。Otsu 使用 ResNet50 模型的二值化方法在 ISIC2017 数据集上的准确率为 91.9%。