CheXpert

Andrew Koh, Alan Tan, Tang Zhe Jun

​

Group 17

1

Outline

• Project Description
• Data acquisition
• Data exploration
• Pre-processing
• Proposed deep learning solutions
• Results
• Conclusion

2

CheXpert

Task

Automatic multi-label classification of chest x-ray images into any of the 5 pathologies (atelectasis, cardiomegaly, consolidation, edema, pleural effusion)

Dataset

223414 train, 234 validation images

Meta-data csv file containing image paths and multi-label binary encoding of pathologies

​

​

Atelectasis

Cardiomegaly

Consolidation

Edema, Pleural Effusion

Pleural Effusion, Consolidation, Edema

3

4

Data Insights

​

More frontal over lateral images

​

Imbalanced dataset

​

~40% images contain 2 or more labels

​

Mean and median age is in range of 63

​

Males are more likely to be contracted!

​

Raw images needs equalization

5

Important Contour Landmarks

#1 - Aortic arch

#2 - Cardiodiaphragmatic angles

#3 - Costophrenic angles

#4 - Hila and other vascular structures

#5 - Rib cage contour

​

​

Helpful to isolate image to only the lung field region

​

Prevent model from learning spurious signals from other parts of the images such as exposure and texts

Simple algorithm to trim images:

​

# trim top of image

top = 0, topMean = np.mean(image[top, :])

while topMean < 60:

top += 1

topMean = np.mean(image[top, :])

​

# do the same for bottom, left and right...

​

return image[top:bottom, left:right]

​

​

~40% of images were badly cropped!

(black borders, off center, etc…)

6

Mean image over 10 bad images�(bad images = off center or contains black borders)

7

=

+

+

+

+

+

8

Batch Size: 16

Epochs: 20

Learning rate=0.001

Input Image size = 320

Optimizer = ADAM

Loss = BCEwithLogitsLoss()

​

Ground truth Labels



Currently:�Impute all unmentioned values to 0

Impute all uncertain values to 1

10

Training Process, Further Explorations & Results

​

• Parallel training on multiple GPUs
• Used Tensorboard to monitor training progress

Models Used:

• ResNet 18
• ResNet 152
• DenseNet 121
• DenseNet 201
• SqueezeNet
• ATP DenseNet 121 *
• Ensemble Stacking *
• Distilled Model - SqueezeNet *

11

Further Exploration

Ensemble

• Combine outputs of top 3 best performing model to form a new model
• Top 3 Models:
• ATP DenseNet 121
• ResNet 18
• DenseNet 121
• Results
• Ensembled model performs worse than ATP DenseNet 121

12

Further Exploration - Knowledge Distillation

• Knowledge Distillation
• Approximate complex models with simpler models
• Allow simpler model to learn from complex model directly
• Initially prepared for ensembled network
• ATP DenseNet121 → SqueezeNet
• Training Phase:
• Match simpler model logits to complex model logits and output
• Loss = ⍺․MSELoss(S_{zero mean},T_{zero mean}) + (1 - ⍺).BCEWithLogitLoss(S, L)
• Set ⍺ = 0.75 after hyperparameter search

13

Overall Results

14

• Unable to do a direct comparison with baseline model ∵ Baseline uses high res images
• But when baseline model (DenseNet121) is trained on low res images, ours perform better
• Using KD, performance of squeezenet improves significantly
• Conclusion
• (Our) ATP DenseNet121 performs better as a whole
• (Our) Distilled SqueezeNet performance matches ResNet152 at only 1.25% size

Summary and main contributions

• Frame crop to preprocess and remove noise from the images
• New architecture: ATP DenseNet121 that improves performance on downsampled images only with a small computational and memory cost (<1MB).
• Knowledge Distillation: Distilled SqueezeNet from ATP DenseNet 121 performance matches ResNet152 at only 1.25% size

​

Full code repository can be found at:�https://github.com/alanwuha/ce7454-grp17

End

References

[1] Wei Dai, Nanqing Dong, Zeya Wang, Xiaodan Liang, Hao Zhang, and Eric P Xing. Scan: Structure correcting adversarial network for organ segmentation in chest x-rays. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, pages 263–273. Springer, 2018.

[2] Hinton, G., Vinyals, O. and Dean, J. (2019). Distilling the Knowledge in a Neural Network. [online] arXiv.org. Available at: http://arxiv.org/abs/1503.02531 [Accessed 13 Nov. 2019].

​

​