16. 深度学习自动化
Input
image
X
Output
variable
Y
Learning from {(Xi , Yi)}
arg minW Si Loss( Yi, f(Xi; W) ) + Reg(W)
Algorithm:
Deep network
Y = f(X; W)
GAN
目标明确
Self-, semi-, or
weakly supervised
标注高效
Neural architecture search
(NAS) or meta-learning
结构合适
New
representation
表达新颖
17. Universal lesion detection 通用肿瘤检测
FPs per
image
@0.5
Fast RCNN
+Bmap
63.96
(+6.79)
FPN-3DCE
+Bmap
69.09
(+4.84)
MVP-Net
+Bmap
73.32
(+3.31)
Bounding Map (BMap)
Li et al. Bounding map for universal lesion detection. MICCAI 2020.
18. Self-supervised learning (SSL) 自监督
a plethora of
unlabelled images
pre-trained model
a few
labelled images
target model
target
task
proxy or self-
supervised
task
Network
learning
Network
learning
19. Self supervised learning (SSL)自监督
Zhu et al. Rubik's cube+: A self-supervised feature learning framework for 3D medical image analysis. Medical Image Analysis 2020.
CT cerebral hemorrhage detection
Train-
from-
scratch
Rubik's
cube+
52.7% 79.5%
MR brain lesion seg. (Dice%)
Train-
from-
scratch
Rubik's
cube+
66.2 72.5
魔方复原
20. :fusing multiple proxy tasks
Aggregative SSL 融合自监督
CT cerebral hemorrhage detection
Train-
from-
scratch
SC-AGGL +
Cube
81.1% 90.2%
STL10
SimCLR 2D Rot
73.1% 70.1%
SRC ALL THREE
68.4% 79.4%
• Fusing multiple proxy tasks
• Less similar, more complementary
Zhu et al. Aggregative self-supervised feature learning. arXiv:2012.07477
21. Partially supervised learning (PSL) 部分监督
Method
#
samples
Avg.
Dice
Multiclass (FSL) 30 .874
Binary class
(FSL)
L126/
S41/
P281/
K210
.851
Multiclass (PSL) 688 .931
Shi et al. Marginal loss and exclusion loss for partially supervised multi-organ segmentation. arXiv:2007.03868. MedIA major revision.
全部
肝 胰
肾
脾
p0
p1
p3
p2
p4 p5
p1
p0+p2+p3+p4+p5
p2
p0+p1+p3+p4+p5
p3
p0+p1+p2+p4+p5
p0+p1+p2+p2
p4
p5
22. Label-free COVID-19 segmentation 无标注
Dice% Corona-
case
Radio-
pedia
UESTC
Anomaly 29.7 32.3 27.2
USL 69.8 59.3 61.4
Inf-Net 66.9 67.8 63.9
Yao et al. Label-free segmentation of COVID-19 lesions in lung CT. IEEE TMI. (major revision)
Synthetic “lesion” generator
24. Universal U-Net
• 6 heterogeneous organ segmentation tasks
• Similar Dice with 1% parameters
• Ability to adapt to new domain
Huang et al. 3D U2-Net: A 3D universal U-net for multi-domain medical image segmentation. MICCAI 2019.
通用+差异化表征
25. Self-inverse network 自逆网络
F = F-1
Shen et al. Learning a self-inverse network for bidirectional MRI image synthesis. ISBI 2020.
Y=F(X)
X=F-1(Y)
↑ 2.7-3.2dB
29. Chest x-ray decomposition and diagnosis
胸片分解与诊断
DNN
decomposition
diagnosis
SOTA accuracy in
predicting 11 out of 14
common lung diseases
based on Chest-xray14
Li et al. High-resolution chest x-ray bone suppression using unpaired CT structural priors, IEEE TMI 2020.
30. Artifact disentanglement network (ADN)
无配对伪影去除
Liao et al. ADN: Artifact disentanglement network for unsupervised metal artifact reduction. IEEE TMI 2020.
分离内容与伪影
PSNR: over 3dB
better than SOTA
31.
32. Medical slice synthesis
Peng et al. SAINT: Spatially aware interpolation network for medical slice synthesis. CVPR 2020.
Spatial
resolution
34. 最新综述
• S. Kevin Zhou, H. Greenspan, C.
Davatzikos, J.S. Duncan, B. van Ginneken,
A. Madabhushi, J.L. Prince, D. Rueckert,
R.M. Summers, A review of deep learning
in medical imaging: Image traits,
technology trends, case studies with
progress highlights, and future promises.
Proceedings of IEEE (minor revision).
arXiv 2008.09104.
36. Advisory Board
1. IMAGING I/O – Stephen Aylward
2. DATA DIVERSITY – Brad Genereaux
3. REPRODUCIBILITY – Lena Maier-Hein
4. TRANSFORMATIONS – Jorge Cordoso
5. FEDERATED LEARNING – Jayashree Kalapathy
6. ADVANCED RESEARCH – Paul Jaeger
7. COMMUNITY ADOPTION – Prerna Dogra
Working Groups
Liaison with the community:
Recommend policies and priorities to development team
Network of AI thought leaders