Due to melanoma is one of the skin cancers with the highest mortality rate and have a large amount of data during the collection and diagnosis, there is an urgent need to improve the diagnostic efficiency and accuracy. However, there remain problems in analyzing medical big data for skin lesion application, such as the intra-class variation and inter-class similarity in skin lesion images and the lacks of ability to focus on the lesion area affecting the classification results of the model. To address these dilemmas, in this paper, we proposed a novel machine learning-based approach that builds on top of DenseNet. It combines the attention mechanism and large margin loss to enhance the classification accuracy in terms of intra-class compactness and inter-class separability. We evaluated our model on ISIC 2017 (International Skin Imaging Collaboration) dataset, which has achieved 92% of Mean AUC. The experimental results show the effectiveness of our solution outperforms the state-of-the-art significantly in classify skin lesion and can accurately classify malignant melanoma on medical images.

由于黑素瘤是死亡率最高的皮肤癌之一，并且在收集和诊断过程中具有大量数据，因此迫切需要提高诊断效率和准确性。但是，在分析用于皮肤病变的医学大数据时仍然存在问题，例如皮肤病变图像中的类内变异和类间相似性，以及缺乏关注影响模型分类结果的病变区域的能力。为了解决这些难题，在本文中，我们提出了一种基于DenseNet的新颖的基于机器学习的方法。它结合了注意力机制和较大的余量损失，从而在类内紧凑性和类间可分离性方面提高了分类的准确性。我们在ISIC 2017（国际皮肤影像协作组织）数据集上评估了我们的模型，该数据集已达到平均AUC的92％。实验结果表明，在对皮肤病变进行分类方面，我们的解决方案的有效性明显优于最新技术，并且可以在医学图像上准确分类恶性黑色素瘤。