China has a centrally funded Healthcare system that covers diagnosis and treatment of tuberculosis. Hence, a misdiagnosis or late-diagnosis of a positive TB case coupled with the high population density will have a significant socio-economic impact on China's healthcare system. While it is true that China has shown a decline in the number of TB cases reported each year, the figures are still quite high.
The following graph shows the distribution of TB cases forecasted in China in 2018,
Journal of Infection and Public Health Volume 11, Issue 5, September–October 2018
To counter this problem my team developed an image classification algorithm to triage tuberculosis patients with an accuracy of 82.5% reducing patient safety issues and patient costs, while having a significant impact on healthcare delivery by providing clinical decision support.
The Dataset: We got the dataset from the Lister Hill National Center for Biomedical Communications, National Library of Medicine. It was collected by Shenzhen No. 3 Hospital in Shenzhen, Guangdong, China. It consists of 326 normal x-rays and 336 abnormal x-rays showing different variations of TB. It also contained information such as gender, age and type of TB for each x-ray.
Data preprocessing: We used histogram equalisation on each x-ray to improve contrast.
We further applied the following data augmentation techniques:
Basic data augmentation techniques to prevent overfitting.
Each training image, was flipped 0, 90, 180, or 270 degrees.
Each image was either flipped left to right or not.
Each image had some random small amount of gaussian noise added to each pixel value.
Deep Learning Model: Using the TensorFlow library and the VGG16 convolutional neural network architecture (built on the ImageNet database) which consists of 13 convolutional layers, 5 max pooling layers, and 3 fully connected neural networks, we created a deep learning model to classify the training set images.
We ran the model on a p2.8xlarge GPU instance via the Amazon Elastic Compute Cloud service. To achieve better accuracy during the validation phase, we tuned the hyper-parameters and ran the model for 10 epochs.
Result: Finally, we achieved an accuracy of 82.5% and a loss of 2.76 in the validation phase.
References:
Dataset: Tuberculosis Chest X-ray Image Data Sets, https://lhncbc.nlm.nih.gov/publication/pub9931
Python Notebook with the code https://github.com/DamienTheCat/CXR_Classification_VGG-16
Efficient Deep Network Architectures for Fast Chest X-Ray Tuberculosis Screening and Visualization, F. Pasa, V. Golkov, F. Pfeiffer, D. Cremers, and D. Pfeiffer, Published online 2019 Apr 18, NCBI NLM NIH.
Forecasting the incidence of tuberculosis in China using the seasonal auto-regressive integrated moving average (SARIMA) model, Qiang Maoa, Kai Zhanga, Wu Yanb, Chaonan Chenga, Journal of Infection and Public Health, Volume 11, Issue 5, September–October 2018.
Engineering for good - detecting pneumonia in X-Ray images, Github: https://github.com/unit8co/amld-workshop-pneumonia
Deep Learning for abnormality detection in Chest X-Ray images http://cs231n.stanford.edu/reports/2017/pdfs/527.pdf
Computer-aided Diagnosis, https://en.wikipedia.org/wiki/Computer-aided_diagnosis#A_Brief_History
Comments