Solutions to Limited Annotation Problems of Deep Learning in Medical Image Segmentation

Image segmentation holds broad applications in medical image analysis, providing crucial support to doctors in both automatic diagnosis and computer-assisted interventions. The heterogeneity observed across various medical image datasets necessitates the training of task-specific segmentation models. However, effectively supervising the training of deep learning segmentation models typically demands dense label masks, a requirement that becomes challenging due to the constraints posed by privacy and cost issues in collecting large-scale medical datasets. These challenges collectively give rise to the limited annotations problems in medical image segmentation. In this dissertation, we address the challenges posed by annotation deficiencies through a comprehensive exploration of various strategies. Firstly, we employ self-supervised learning to extract information from unlabeled data, presenting a tailored self-supervised method designed specifically for convolutional neural networks and 3D Vision Transformers. Secondly, our attention shifts to domain adaptation problems, leveraging images with similar content but in different modalities. We introduce the use of contrastive loss as a shape constraint in our image translation framework, resulting in both improved performance and enhanced training robustness. Thirdly, we incorporate diffusion models for data augmentation, expanding datasets with generated image-label pairs. Lastly, we explore to extract segmentation masks from image-level annotations alone. We propose a multi-task training framework for ECG abnormal beats localization and a conditional diffusion-based algorithm for tumor detection.

图像分割在医学影像分析中具有广泛应用场景，可为医师的自动诊断与计算机辅助干预提供关键支撑。各类医学影像数据集存在的异质性，故而需要针对具体任务训练专用分割模型。然而，有效监督深度学习分割模型的训练通常需要稠密标注掩码，但由于大规模医学影像数据集的采集受隐私与成本因素制约，满足这一要求颇具挑战。上述各类挑战共同导致了医学影像分割领域的标注不足问题。针对标注不足引发的各项挑战，本论文通过全面探索多种策略开展研究。其一，我们利用自监督学习从无标注数据中提取信息，提出了专为卷积神经网络（Convolutional Neural Networks）与3D视觉Transformer（3D Vision Transformers）设计的定制化自监督方法。其二，我们将研究重心转向域自适应问题，利用内容相似但模态不同的影像数据；我们在图像转换框架中引入对比损失作为形状约束，既提升了模型性能，也增强了训练鲁棒性。其三，我们引入扩散模型（Diffusion Models）用于数据增强，通过生成影像-标注对扩充数据集规模。其四，我们探索仅依靠图像级标注提取分割掩码的方法，提出了用于心电图（Electrocardiogram, ECG）异常心搏定位的多任务训练框架，以及用于肿瘤检测的基于条件扩散的算法。