Wang Jiaxin, Zhu Weifang, Xiang Dehui, Chen Xinjian, Peng Tao, Peng Qing, Wang Meng, Shi Fei
MIPAV Lab, School of Electronics and Information Engineering, Soochow University, Suzhou, China.
The State Key Laboratory of Radiation Medicine and Protection, Soochow University, Suzhou, China.
Med Phys. 2025 Sep;52(9):e18102. doi: 10.1002/mp.18102.
Deep learning-based segmentation methods for optical coherence tomography (OCT) have demonstrated outstanding performance. However, the stochastic distribution of training data and the inherent limitations of deep neural networks introduce uncertainty into the segmentation process. Accurately estimating this uncertainty is essential for generating reliable confidence assessments and improving model predictions.
To address these challenges, we propose a novel uncertainty-guided cross-layer fusion network (UGCFNet) for retinal OCT segmentation. UGCFNet integrates uncertainty quantification into the training process of deep neural networks and leverages this uncertainty to enhance segmentation accuracy.
Our model employs an encoder-decoder architecture that quantitatively assesses uncertainty at multiple stages, directing the network's focus toward regions with higher uncertainty. By facilitating cross-layer feature fusion, UGCFNet enhances the comprehensive understanding of both semantic information and morphological details. Additionally, we incorporate an improved Bayesian neural network loss function alongside an uncertainty-aware loss function, enabling the network to effectively utilize these mechanisms for better uncertainty modeling.
We conducted extensive experiments on the publicly available AI-Challenger and OIMHS OCT segmentation datasets. The training, validation, and testing sets of the AI-Challenger dataset are comprised of 32, 8, and 43 OCT volumes, yielding a total of 4096, 1024, and 5504 B-scans, respectively. The training, validation, and testing sets of the OIMHS dataset consist of 100, 25, and 25 OCT volumes, resulting in 2,310, 798, and 751 B-scans, respectively. The results demonstrate that UGCFNet achieves state-of-the-art performance, with average Dice similarity coefficients of 79.47% and 93.22% on the respective datasets.
Our proposed UGCFNet significantly advances retinal OCT segmentation by integrating uncertainty guidance and cross-level feature fusion, offering more reliable and accurate segmentation outcomes.
基于深度学习的光学相干断层扫描（OCT）分割方法已展现出卓越性能。然而，训练数据的随机分布以及深度神经网络的固有局限性给分割过程带来了不确定性。准确估计这种不确定性对于生成可靠的置信度评估和改进模型预测至关重要。
为应对这些挑战，我们提出了一种用于视网膜OCT分割的新型不确定性引导跨层融合网络（UGCFNet）。UGCFNet将不确定性量化集成到深度神经网络的训练过程中，并利用这种不确定性来提高分割精度。
我们的模型采用编码器 - 解码器架构，在多个阶段对不确定性进行定量评估，引导网络将注意力集中在不确定性较高的区域。通过促进跨层特征融合，UGCFNet增强了对语义信息和形态细节的全面理解。此外，我们结合了改进的贝叶斯神经网络损失函数和不确定性感知损失函数，使网络能够有效利用这些机制进行更好的不确定性建模。
我们在公开可用的AI - Challenger和OIMHS OCT分割数据集上进行了广泛实验。AI - Challenger数据集的训练集、验证集和测试集分别由32、8和43个OCT体积组成，分别产生总共4096、1024和5504个B扫描。OIMHS数据集的训练集、验证集和测试集分别由100、25和25个OCT体积组成，分别产生2310、798和751个B扫描。结果表明，UGCFNet实现了领先的性能，在相应数据集上的平均骰子相似系数分别为79.47%和93.22%。
我们提出的UGCFNet通过集成不确定性引导和跨层特征融合，显著推进了视网膜OCT分割，提供了更可靠、准确的分割结果。
