基于扩散模型确定性病变表征的眼底病灶分割方法

Retinal Lesion Segmentation with Retinopathy Guidance Deterministic Representation from Diffusion Models

获取全面的眼底病变分割图是开发自动化、可解释的视网膜病症诊断工具的关键步骤。然而,眼底图像上病变的多样性及复杂性,导致了精确标注的稀缺,限制了传统监督学习方法的发展及应用。近期研究表明,表征学习通过从大规模未标注数据中预训练强大的图像表征提取模型,在下游任务中仅需少量标注数据即可取得优异的性能表现。本研究提出了一种新颖的基于去噪扩散概率模型的表征学习分割框架。这一框架的目标是通过生成式建模,更精准地捕捉医学图像中的局部和微妙变化,为眼底图像的病灶分割提供精确的特征表示。采用未标记的眼底图像来学习预定义的马尔科夫扩散的逆过程,从而为从眼底图像中提取像素级表征奠定基础。此外,考虑到视网膜病变的严重性和病灶的相关性,引入一个病变分级网络,以指导逆扩散过程,增强与病灶紧密相关的表征能力。这些经过引导的表征作为眼底图像内在语义信息的存储库,为下游视网膜分割任务提供坚实的图像像素级表征基础。在多个眼底图像数据集上的实验中,所提出的方法在视杯和视盘分割任务上仅使用50个样本取得了0.872和0.877的平均Dice系数。在糖尿病性视网膜病变病灶分割中,平均Dice系数为0.664,而在年龄相关性黄斑变性病灶分割任务中,模型达到了0.513的平均Dice系数。研究结果证明了扩散模型所学习到的表征在多种复杂眼底病变分割任务上的通用性和有效性。

Acquiring a comprehensive segmentation map of retinal lesions is a crucial step in developing an automated,interpretable diagnostic tool for retinopathy.However,the inherent diversity and complexity of retinal lesions,coupled with the high cost of precise annotation,pose substantial challenges to traditional supervised learning approaches.Recent advances suggest that representation learning can mitigate the reliance on extensive annotated data by pre-training robust image representation models on large-scale unlabeled datasets.In this study,we introduced an innovative representation learning framework based on denoising Diffusion Probabilistic Models(DDPM),specifically tailored to capture the subtle and localized variations in medical imagery,thereby providing precise feature representations for the segmentation of retinal lesions.Utilizing unlabeled fundus images,our approach learnt the reverse process of Markov diffusion,establishing a foundation for extracting pixel-level representations.A retinal lesion grading classifier,informed by domain knowledge of retinopathy severity and lesion correlation,was implemented to guide the reverse diffusion process to enhance representations pertinent to lesions.The guided representations served as a repository of intrinsic semantic information,offering robust image representations for downstream retinal segmentation tasks.In experiments on multiple fundus image datasets,our method achieved average Dice coefficients of 0.872 for optic cup and 0.877 for optic disc segmentation with only 50 samples.For diabetic retinopathy lesions,it reached a Dice coefficient of 0.664,and for age-related macular degeneration lesions,0.513,demonstrating diffusionbased representation's generality and effectiveness across various complex retinal conditions.