基于半监督学习的糖尿病视网膜病变自动评级和分割

Automated Diabetic Retinopathy Grading and Segmentation Based on Semi-Supervised Learning

摘要: 糖尿病视网膜病变(DR)是导致永久性失明的主要原因之一，尤其是在中老年人中。早期发现DR对于预防视力丧失至关重要。疾病严重程度分级可以看作是一个分类问题，只需要图像级的标注，而病变分割则需要更强的像素级标注。然而逐像素数据标注非常耗时，需要专家。本文提出了一种协同学习的方法来联结优化DR评级和病灶检测。首先利于小部分逐像素数据标注数据训练一个多病变标签语义分割模型，接着基于最初对大量图像级注释数据预测的病变图，设计了一个疾病注意力评级模型。同时，疾病注意力模型可以利用特定类的信息对病灶图进行细化，以半监督的方式对分割模型进行微调。生成对抗性架构也被应用在训练过程中。接着在三个公共数据集上对本文的方法进行验证，并与一些先进的模型进行对比，取得较好的性能。

Abstract: Diabetic retinopathy (DR) is one of the leading causes of permanent blindness, especially among middle-aged and elderly individuals. Early detection of DR is crucial for preventing vision loss. The severity grading of the disease can be seen as a classification problem, requiring only image-level annotations, while lesion segmentation requires more detailed pixel-level annotations. However, pixel-wise data annotation is time-consuming and requires expertise. In this paper, we propose a collaborative learning approach to integrate the optimization of DR grading and lesion detection. Firstly, we train a multi-label semantic segmentation model using a small subset of pixel-wise annotated data, then we design a disease attention grading model based on the initial predictions of lesion maps from a large amount of image-level annotated data. Meanwhile, the disease attention model can refine lesion maps using specific class information to fine-tune the segmentation model in a semi-supervised manner. Generative adversarial architectures are also applied in the training process. We then validate our approach on three public datasets and compare it with some state-of-the-art models, achieving better performance.