Dermoscopic Image Classification with Neural Style Transfer
提出将神经风格迁移作为皮肤镜图像预处理步骤，通过风格迁移和张量分解提取特征，在ISIC数据集上分类准确率提升超10%，并可与预训练CNN模型竞争。
AbstractSkin cancer, the most commonly found human malignancy, is primarily diagnosed visually via dermoscopic analysis, biopsy and histopathological examination. However, unlike other types of cancer, automated image classification of skin lesions is deemed to be more challenging due to the irregularity and variability in the lesions’ appearances. In this work, we propose an adaptation of the Neural Style Transfer (NST) as a novel image preprocessing step for skin lesion classification problems. We represent each dermoscopic image as a style image and transfer the style of the lesion onto a homogeneous content image. This transfers the main variability of each lesion onto the same localized region, which allows us to integrate the generated images together and extract low-rank latent embeddings via tensor decomposition. We evaluated the performance of our model on competition datasets collected and preprocessed from the International Skin Imaging Collaboration (ISIC) database. We show that the classification performance based on the extracted tensor features using the style-transferred images significantly outperforms that of the raw images by more than 10%, and is also competitive with well-studied, pretrained CNN models using transfer learning. Additionally, the tensor decomposition also affords clinical interpretations and insights by examining the images which correspond to the largest loadings in the top style embedding features as identified by the common supervised learning models. Supplementary materials for this article are available online.KEYWORDS: CNNMedical image preprocessingMelanoma classificationTensor decomposition Supplementary MaterialsThe supplementary material contains the detailed implementation of the U-net model, ABCD rule, and VGG networks for classification. In addition, the data set along with the details of how to reproduce the results in the manuscript is also provided.AcknowledgmentsThe authors would like to acknowledge the editor, associate editor, and anonymous referees for their critical and insightful comments in improving this article.Additional informationFundingThis work was supported by the National Science Foundation (NSF) under Grant DMS-1952406 and Grant DMS-1821198.