Pneumonia detection using transfer learning and explainable AI

Project Title

Pneumonia detection using transfer learning and explainable AI

Project Area of Specialization Artificial IntelligenceProject Summary PROJECT SUMMARY

Pneumonia is a principal cause of death in infants and is an infectious disease in general that affects the lungs. Although deep learning methods have been effective for a variety of diagnostic tasks and have beaten some human experts, the black-box nature of algorithms has restricted radiologists to put trust in them. The idea is to build an automated system that distinguishes chest X-rays of patients with Pneumonia while showing the features that influence the decision of the model, making it more trustable. The main hurdle in training such a system is the scarcity of datasets in medical imaging. We will explore different techniques to cater this problem. Furthermore, we will be applying two to three transfer learning techniques for training the model as they tend to work best with smaller datasets. The trained models will be assessed on their sensitivity, Explainable Artificial Intelligence(XAI) methods will then be applied to the models giving the best performance.

Project Objectives PROJECT OBJECTIVES

The objective of this research is as follows:

• Firstly to deal with the class imbalance problem in our data set “CX8” between two classes “Pneumonia” and “No Findings”. 

• Secondly, comparing a few transfer learning approaches and analyzing which one of them gives robust results in terms of sensitivity for our dataset. 

• Lastly, justifying the result using explainable artificial intelligence (XAI) to gain trust of the medical practitioners. Moreover, developing an interactive Graphical User Interface (GUI) to facilitate them.

Project Implementation Method PROJECT IMPLEMENTATION METHODS

The Explainable pneumonia diagnosis system will be using CXR8[1] ImageNet dataset by NIH Clinical Center. The more data one provides to the ML/DL model, the quicker that model can learn and upgrade and when it comes to medical image analysis we need to have a huge amount of data for the detection of even the smallest abnormality and for achieving accuracy. In the pre-processing phase we will cater the imbalance problem by trying a combination of under sampling methods like evolutionary undersampling and oversampling techniques including Generative Adversarial Network(GAN)[2] and data augmentation. We look forward to using the transfer learning approach to train the model. We will be applying three pre-trained models Alexnet[3 ], Resnet[4] & Exception[5]. The performance of these models will be assessed on the basis of sensitivity. The model giving the best results will be used for Explainable Artificial Intelligence(XAI)[6 ]. The methods we will be using for XAI are the attribution maps (heat maps) generated by Class Activation Map[7] (CAM) technique as it categorizes the important parts/pixels of an image along with this we are going to apply perturbation based and back propagation based methods as per compatibility for better explanation and interpretability of the model.

REFERENCES:

1. X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases,” CoRR, vol. abs/1705.02315, 2017. [Online]. Available: http://arxiv.org/abs/1705.02315

2.  I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014.

3. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” in Advances in Neural Information Processing Systems, F. Pereira, C. J. C. Burges, L. Bottou, and K. Q. Weinberger, Eds., vol. 25. Curran Associates, Inc., 2012. [Online]. Available: https://proceedings.neurips.cc/paper/2012/ file/c399862d3b9d6b76c8436e924a68c45b-Paper.pdf 

4. K. He, X. Zhalng, S. Ren, and J. Sun, “Deep residual learning for image recognition,” Dec. 2015.

5. F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” CoRR, vol. abs/1610.02357, 2016. [Online]. Available: http://arxiv.org/abs/1610.02357

6. Shaw-Hwa Lo, Yiqiao Yin. A NOVEL INTERACTION-BASED METHODOLOGY TOWARDS EXPLAINABLE AI WITH BETTER UNDERSTANDING OF PNEUMONIA CHEST X-RAY IMAGES. ArXiv. Volume abs/2104.12672, November 22, 2021

7. Hyungsik Jung Youngrock Oh. Towards Better Explanations of Class Activation Mapping. Axriv. Vol 3, Mon, 27 Sep 2021.

Benefits of the Project BENEFITS OF THE PROJECT 

Pneumonia is one of the biggest threats to human life all over the world. Early diagnosis of pneumonia is critical for identifying the best course of treatment and further preventing the disease from posing a life-threatening hazard to the patient, thus if the gap in diagnosing various types of pneumonia is not bridged with robust automated techniques of chest disease identification, the healthcare sector may be forced to face undesirable situations. For this matter we proposed a Computer-Aided Diagnosis (CAD) technique that assists doctors to detect and interpret various types of abnormalities in medical imaging, to diagnose and analyze diseases accurately and promptly.

The Explainability methods (XAI) that are to be used in our project, assist to visualize the features most responsible for a model's decision, thus making it interpretable just as simple models are so that these systems are transparent, understandable, and explainable to gain the trust of physicians, regulators as well as the patients. And to further facilitate the physicians, medical practioners and regulators we look forward to develop an application with user friendly graphical user interface in future. 

Technical Details of Final Deliverable TECHNICAL DETAILS OF FINAL DELIVERABLES

An automated Computer Aided Diagnosis (CAD) system was proposed in this study to aid medical practitioners for pneumonia diagnosis. The system uses a transfer learning approach to classify X-ray images into two classes “Pneumonia” and "No Finding". We further aim to develop an interactive Graphical User Interface (GUI) in order to facilitate medical practitioners as they'll be able to add patients chest X-Ray image to the system and generate the results, which will depict the probabilities of a patient having "Pneumonia" and "No Pneumonia". Furthermore, heat maps will be generated of the scanned X-Ray image which highlights the input pixels against the classification output for providing an explainable (XAI) solution which can be easily understood by humans as to how and on what basis the model is generating the result so as to gain the trust of physicians, regulators as well as the patients.

Final Deliverable of the Project Software SystemCore Industry MedicalOther Industries IT , Health Core Technology Artificial Intelligence(AI)Other Technologies OthersSustainable Development Goals Good Health and Well-Being for PeopleRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	36590
Google storage	Equipment	10	188	1880
Google Colaboratory	Equipment	10	1871	18710
Traveling	Miscellaneous	8	500	4000
Stationer	Miscellaneous	6	1000	6000
Ardino	Equipment	2	3000	6000