一、图像多模态
这里指的是有一些具有多个模态的医学图像,例如 MRI,或者不同成像方式得到的图像,比如 CT-MR,PET-MR 组成的多模态医学数据。
多模态融合方式包括在输入的融合,layer-level 的融合,以及在决策端 or 输出的融合。
1. Input-level fusion
多数模型采用的方法:UNet,nnUNet, CNN+ViT,UNETR_v, Swin UNETR。
1.1 Convolution-based:
几个模态拼接在一起输
1.2 Transformer-based:
Multi-modal medical Transformers: A meta-analysis for medical image segmentation in oncology,2023
i. tokenization:每个模态单独tokenize,或者多模态拼接起来tokenize
ii. linear projection:共用一个,或者每个模态单独一个 projection
iii. before transformer block:每个模态在对应位置加权求和,或者concat
2. Layer-level fusion
2.1 卷积网络
i. Modality-Aware Mutual Learning for Multi-modal Medical Image Segmentation,2021
两个模态单独分支和 concat 一起的分支融合
ii. Cross-Modal Prostate Cancer Segmentation via Self-Attention Distillation,2021
两个模态分别输入在中间层融合
iii. Cross-modality deep feature learning for brain tumor segmentation,2021
基于 GAN 的方法,先训练模态间的生成器,然后模态融合做分割
iv. Feature-enhanced generation and multi-modality fusion based deep neural network for brain tumor segmentation with missing MR modalities,2021
生成缺失模态,然后分割
2.2 Transformer
i. 模态间的 cross-attention:
把所有 block 里都换成 cross-attention,或者结合 cross-attention 和 self-attention
3. Decision-level fusion
决策层上的融合时,每个模态图像都用作单个分割网络的单个输入。单一网络可以更好地利用相应模态的独特信息。然后将整合各个网络的输出以获得最终的分割结果。
决策层的融合策略包括平均和投票。平均策略通常对各个网络的置信度进行平均。通过为每个体素分配最高置信度来获得最终的分割。对于多数投票策略,体素的最终标签取决于各个网络的大多数标签。
决策层融合的缺点是需要更多内存,因为需要训练更多的参数。
4. 结论
Multi-modal medical Transformers: A meta-analysis for medical image segmentation in oncology,2023
i. ViT+CNN混合模型表现最好,但是取决于任务,随着模态数量增加,操作数增加
ii. nnUNet 在不同任务上表现稳定,well-traine nnUNet在中小型数据集的多模态分割任务上足够了
iii. 建议使用nnUNet pipline
STU-Net: Scalable and Transferable Medical Image Segmentation Models Empowered by Large-Scale Supervised Pre-training, 2023
i. nnUNet比transformer-based 网络在多种任务上更稳定
ii. nnUNet 的up-scale版本,在大数据集上预训练后在多模态数据集上微调,效果比nnUNet好
5. paper list
- Multi-Modal CO-learning for Live Lesion Segmentation on PET-CT Image, TMI 2021
- Disentangle domain features for cross-modality cardiac image segmentation, MIA 2021
- United adversarial learning for liver tumor segmentation and detection of multi-modality non-contrast MRI, MIA 2021
- Multi-phase and Multi-level Selective Feature Fusion for Automated Pancreas Segmentation from CT lmages, MICCAI 2020
- Semi-Supervised Unpaired Multi-Modal Learning for Label-Efficient Medical lmage Segmentation, MICCAI 2021
- Semantic Consistent Unsupervised Domain Adaptation for Cross-Modality Medical lmage Segmentation, MICCAI 2021
- Modality-Aware Mutual Learning for Multi-modal Medical lmage Segmentation, MICCAI 2021
- Generalizable Cross-modality Medical Image Segmentation via Style Augmentation and Dual Normalization, CVPR 2022
- SDC-UDA, cvpr2023
网友评论