摘要：

结直肠内窥镜图像在设备之间的特征分布不同，导致训练模型在新设备上的分割性能降低。为缓解模型对新设备的适应性问题，提出一种基于增量学习的微调方法，以及一种改进的结直肠息肉分割网络CPSegNet。增量学习方法包含预训练和新设备微调2个阶段，预训练使用旧设备的数据对息肉分割网络进行充分训练，微调阶段同时使用新旧设备样本进行训练，并通过采样率和正则化损失函数防止出现灾难性遗忘现象。CPSegNet采用MiT的预训练模型作为骨干网络，多层感知机(MLP)作为解码模块，不确定区域注意力(URA)作为细化模块，对边界模糊区域进行优化。为了验证学习策略对新设备的适应能力，采用Kvasir-SEG、CVC-ClinicDB、CVC-300、CVC-ColonDB、Kvasir-Sessile和ETIS-LaribPolypDB共6个数据集进行实验，其中前2个数据集为训练集，其余4个为新设备的模拟数据。以Dice相似系数和交并比(IoU)作为评价指标。实验结果表明，在无增量学习情况下CPSegNet在新设备上的性能优于主流的算法，以Kvasir-SEG作为源域数据集，将较难分割的ETIS-LaribPolypDB作为目标域数据集时，与ColonFormer算法相比的Dice相似系数提升3百分点，以CVC-ClinicDB作为源域数据集时，提升了6百分点，使用增量学习后CPSegNet和主流算法都能在新设备上获得性能提升，同时保持在旧设备上的分割精度。

关键词: 息肉分割, 增量学习, 迁移学习, 少样本学习, 灾难性遗忘

Abstract:

The feature distributions of colorectal endoscopic images differ among devices, reducing the trained model′s segmentation performance on new devices. To alleviate the model′s adaptability to new devices, a fine-tuning method based on incremental learning and an improved colorectal polyp segmentation network called CPSegNet are proposed. The incremental learning method consists of two stages: pre-training and fine-tuning on new devices. Pre-training uses data from an old device to train the polyp segmentation network adequately, and the fine-tuning stage is trained with samples from both old and new devices. This also includes a sampling rate adjustment and a regularization loss function to prevent catastrophic forgetting. CPSegNet adopts a pre-trained MiT model as the backbone network, a Multi-Layer Perceptron (MLP) as the decoding module, and an Uncertainty Region Attention (URA) mechanism as the refinement module to optimize the ambiguous boundary regions. To validate the adaptability of the learning strategy to new devices, experiments are conducted using six datasets: Kvasir-SEG, CVC-ClinicDB, CVC-300, CVC-ColonDB, Kvasir-Sessile, and ETIS-LaribPolypDB; the first two datasets are used as the training set, and the other four are simulated data for new devices. The experimental results, using the Dice similarity coefficient and Intersection over Union (IoU) metrics as evaluation indicators, demonstrate that the performance of CPSegNet on new devices is superior to that of mainstream algorithms without incremental learning, particularly on the challenging ETIS-LaribPolypDB dataset, showing increases of 3 percentage points in the Dice similarity coefficient compared with the ColonFormer algorithm when Kvasir-SEG is used as the source domain dataset. When CVC-ClinicDB is used as the source domain dataset, the Dice similarity coefficient is improved by 6 percentage points. Furthermore, both CPSegNet and mainstream algorithms exhibit performance improvements on new devices after using incremental learning, while maintaining segmentation accuracy on old devices.

Key words: polyp segmentation, incremental learning, transfer learning, few-shot learning, catastrophic forgetting