Download as pptx, pdf, or txt
Download as pptx, pdf, or txt
You are on page 1of 19

Indian Institute of Information Technology, Nagpur

Dept. of Computer Science and Engineering


Presentation-III on “Few Shot Learning for Chest X-rays”

G roup/Team
Members:
Dushyant Singh
BT19CSE108
Name of the
Supervisor:
BT19CSE07 Shivam Kumar
1 Dr. Tausif Diwan

BT19CSE122 Astitva

BT19CSE09 Yogesh Walecha


1
4
Introduction

• Machine Learning, Deep Learning works well if a large amount of labelled data is present.

• But deep learning models fail to work with less data as they will end up overfitting it.

• Few shot learning can be used to classify a new task just by looking at a few labeled
examples.
Dataset used
• Covid-19 Radiography dataset (Covid-19, Normal, Pneumonia)
• NIH Dataset 13 classes (Base classes)
Dataset
Features of Covid-19 X-ray

• Ground Glass Opacities (GGOs):


Hazy areas in the lungs.

• Consolidation:
Appears as white patches in the lungs.

• Peripheral Distribution:
Peripheral distribution of the lung abnormalities, meaning that they are more
commonly seen in the outer areas of the lungs.
Features of Covid-19 X-ray

Arrows depict Hazy areas and white patches on


the outer areas
Architecture of Our Network
• We have used Transfer Learning + Siamese Network
• Siamese network pairs naturally with Transfer Learning
• Pairs of images as input and it predicts if they are similar or not
• Each image is passed through a CNN feature extractor which gives
us the feature vector of the image.
• Then we combine these 2 feature vectors by taking Euclidian
distance and make the prediction.
Transfer Learning
• We are using VGG-16 classifier as our feature extractor.
• We tried some other models such as Xception , Resnets and
Mobilenet but VGG-16 worked out best for us.
• And instead of using random weights as initializer we are
‘imagenet’
weights.
• Because it will help us in capturing high level features and then we
can finetune the model according to our task.
• So, combining transfer learning with Siamese Network we can
increase performance and save a lot of time
Siamese Architecture
Our Hypothesis
• Hard pairs are the pairs which are from same class and have large
Euclidian distance between their feature vectors
• And the pairs belonging to different class but have similar Euclidian
distance.
• As there are N^2 pairs, randomly selecting we won’t be able to
capture the hard pairs.
• So, we find these hard pairs explicitly
• Training on hard pairs makes our model more model more robust
and it forces it to learn the true distinguishing feature
Continued…
• After every 10th epoch we use our trained model till now and
generate hard pairs equal to half our batch size
• For hard pairs we choose pairs from same class having distance
greater than threshold (> 0.7)
• And the pairs from different classes having less than threshold
distance (< 0.3)
Loss and Accuracy after training

A
C
C
L
U
O
R
S
A
S
C
Y

Epochs Epochs
Base model vs ours

• Around 2-3 % improvement over base model.


• Query example Covid-19 image.
• Out of N images only 1 is Covid-19 image.
• N-way, form N- pairs and take least Euclidian distance pair.
• We do this to make our model task harder as now there are 20 pairs with only 1
possible correct answer.
Our Web Application
• Get diagnosis report from uploading chest Xray.

• Generate heat map of effected regions.

• Trained on CheXpert dataset which contains 14 different


classes of radiology conditions
Upload Xray
• The web application allows user to upload chest X-ray image in any
popular image format
After uploading the image user can see the
predicted classes by model
Heat Maps
• User can also generate heat map which highlights the features CNN
has used to predict the condition.
References
1Li, A.; Luo, T.; Lu, Z.; Xiang, T.; Wang, L. Large-scale few-shot learning: Knowledge transfer with class hierarchy. In Proceedings of the 2019 IEEE/CVF Conference on Computer
Vision and Pattern Recognition, Long Beach, CA, USA, 15–20 June 2019; pp. 7205–7213.
2Finn, C.; Abbeel, P.; Levine, S. Model-agnostic meta-learning for fast adaptation of deep networks. In Proceedings of the 34th International Conference on Machine Learning,
Sydney, NSW, Australia, 6–11 August 2017; pp. 1156–1168.
3 Liu, Y.; Zhang, H.; Zhang, W.; Lu, G.; Tian, Q.; Ling, N. Few-Shot Image Classification: Current Status and Research Trends. Electronics 2022, 11, 1752.
4Gui, L.; Wang, Y.-X.; Hebert, M. Few-shot hash learning for image retrieval. In Proceedings of the 2017 IEEE International Conference on Computer Vision Workshops,
Venice, Italy, 22–29 October 2017; pp. 1228–1237
5 Chelsea Finn, Pieter Abbeel, and Sergey Levine. “Model-agnostic meta-learning for fast adaptation of deep networks." ICML 2017.
6Jamal, M.A.; Qi, G.-J. Task agnostic meta-learning for few-shot learning. In Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition,
Long Beach, CA, USA, 15–20 June 2019; pp. 11711–11719.
7Koch, G.; Zemel, R.; Salakhutdinov, R. Siamese neural networks for one-shot image recognition. In Proceedings of the Conference and Workshop on the 32nd International
Conference on Machine Learning, Lille, France, 6–11 July 2015.
8Vinyals, O.; Blundell, C.; Lillicrap, T.; Kavukcuoglu, K.; Wierstra, D. Matching networks for one shot learning. arXiv 2017, arXiv:1606.04080. Available online:
http://arxiv.org/abs/1606.04080 (accessed on 29 December 2017).
9 Shang, H.; Sun, Z.; Yang, W.; Fu, X.; Zheng, H.; Chang, J.; Huang, J. Leveraging other datasets for medical imaging classification: Evaluation of transfer, multi-task
and semi-supervised learning. In Proceedings of the International Conference on Medical Image Computing and Computer-Assisted Intervention, Shenzhen, China, 13–17 October
2019; Springer: Cham, Switzerland, 2019; pp. 431–439.
10Cai, A.; Hu, W.; Zheng, J. Few-Shot Learning for Medical Image Classification. In Proceedings of the International Conference on Artificial Neural Networks, Bratislava,
Slovakia, 15–18 September 2020; Springer: Cham, Switzerland, 2020; pp. 441–452.
11Chen, X.; Yao, L.; Zhou, T.; Dong, J.; Zhang, Y. Momentum contrastive learning for few-shot covid-19 diagnosis from chest ct images. Pattern Recognit. 2021, 113, 107826
12 Jadon, S. COVID-19 detection from scarce chest X-ray image data using few-shot deep learning approach. In Medical Imaging 2021: Imaging Informatics for
Healthcare, Research, and Applications; International Society for Optics and Photonics: Bellingham, WA, USA, 2021.
13 Liu, L.; Zhou, T.; Long, G.; Jiang, J.; Yao, L.; Zhang, C. Prototype Propagation Networks (PPN) for Weakly-supervised Few-shot
Learning on Category Graph. arXiv 2019, arXiv:1905.04042. Available online: http://arxiv.org/abs/1905.04042 (accessed on 2 June
2019).
14Fei-Fei, L.; Fergus, R.; Perona, P. One-shot learning of object categories. IEEE Trans. Pattern Anal. Mach. Intell. 2006, 28, 594–611. [CrossRef]
15Fe-Fei, L.; Fergus; Perona. A Bayesian approach to unsupervised one-shot learning of object categories. In Proceedings of the Ninth IEEE
International Conference on Computer Vision, Nice, France, 13–16 October 2003; Volume 2, pp. 1134–1141.
Thank
You!

You might also like