MICCAI: International Conference on Medical Image Computing and Computer Assisted Intervention
今年非常有幸參加了在深圳舉辦的MICCAI2019，特将部分會議内容總結如下：
Begin: 1998
構成：CVRMed（計算機視覺，虛拟現實和醫學機器人），
MRCAS（醫學機器人和計算機輔助手術）
VBC（生物醫學可視化）
MIC & CAI
參會人員： 2000年 400人左右；
2017年 > 1000人；
2019年 >2300人；
KeyNote:
Keynote 1: AI Medical Imaging in China: Now and Future
Keynote 2 : AI Medical Imaging in China: Now and Future
Keynote 3 : Psychoradiology: The Forefront of Clinical Neuroimaging
Papers Overview:
- 45篇MRI，30篇CT，18篇Ultrasound，7篇OCT，5篇PET
- 21篇surgery
- 12篇CNN，19篇GAN，28篇Generative，93篇 Deep，150篇 Network
- ......
Loss Functions: Cross Entropy Loss + Dice Loss. Pixel-wise Reweighted Dice (PRD) loss. Image-wise Shape Regularization (ISR) loss. KL Divergence loss. Reconstruction loss. Domain loss. Focal loss. Mesh loss + Point loss. Dense CRF loss. ......
Segmentation Metrics: Dice Coefficient (DC) Mean Absolute Surface Distance (MAD) Mean Absolute Error (MAE) Hausdorff Distance (HD) ACC, AUC SE SP ......
Difficulty and Challenges:
- 3D Image: limitation of GPU memory & lack of pre-trained models & difficulty in training.
- The lack of annotated data.
- The image quality: low SNR & poor contrast.
- High class imbalance (mainly background).
- Significant variations in the shape and appearance of object.
- Small target with large shape variations.
- Hard Cases (Unclear boundary, big or small tumors).
- Missing or ambiguous boundaries.
- Inhomogeneous appearance and image quality.
- ......
Some Conclusion:
- In 2019, the U-Net is still state of the art.
- nnU-Net is a strong U-Net implementation that generalizes well to new datasets.
- 3D U-Net is a strong baseline for segmentation.
- Cascaded pipline is an effective way to improve tumor segmentation performance.
- Modified ground truth masks may lead to better model generalization.
- High realism is not equivalent to meaningful features that improve classification performance.
- Jointly learning registration and segmentation is superior over separate learning.
- TL (Transfer Learning) should be applied for performance improvement.
- MTL (Multi-Task Learning) is goog fit for small data size.
- Coarse to fine approach is important.
- Multi-View method.
- Shape and Spatial Priors are important for performance improvement.
- ......
Other Research:
- Prediction of Hemorrhagic Transformation in Acute Stroke (MRI) .
- Color Normalization of H&E (Hematoxylin and Eosin) Stained Images.
- Quality Evaluation for Noisy-Labeled Image Segmentation.
- Early Prediction of ALzheimer's Disease (MRI).
- Consensus Neural Network for Medical Imaging Denoising with Only Noisy Training Samples.
- Weakly supervision with image-level annotation.
- Data Augmented Samples in Deep Learning.
- Domain Adaptation Approach to Classification.
- Curriculum semi-supervised segmentation.
- Multi-View fusion of 4D ultrasound reveals fetal head.
- Image generation.
- Improve labeling efficiency.
- Image enhancement.
- Text to Mask Label.
- Hyper Graph Learning
- Bone Age Assessment
- 3D + 2D
- ......