Building an object localization algorithm to detect visual signals for pneumonia in X-rays.
About 900,000 US citizens get pneumococcal pneumonia each year, and about 50,000 die with the disease every year in the US.
These deaths can be easily prevented by early detection and targeted antibiotic therapy.
Chest X-rays is an effective tool for diagnosing pneumonia, but the diagnostic accuracy is limited.
Determining the extent and location of the pneumonia infection with total accuracy is one of the pressing challenges faced by the radiologists.
Detecting pneumonia in chest X-rays is a challenging task that relies on the availability of expert and experienced radiologists. We needed to build and present a deep learning-based model that can offer a supervised learning approach to the radiologists by detecting the visual signals for pneumonia. The solution needed to leverage deep learning technology to offer an accurate review of a chest radiograph (CXR) and to be integrated with a machine learning model to learn from current prediction and enhance the quality and accuracy of chest X-ray predictions step by step.
Creating comprehensive architecture for a deep neural network to detect lung opacities.
We started working on designing the right and efficient neural networking models like CNN (Convolutional Neural Networks).
To create and test the algorithm we utilized ChestX-ray 14, the largest publicly available chest Xray dataset provided by the NIH clinical center.
The system was designed to leverage CNN models that can segment the given image, using the bounding boxes directly as a mask.
We decided to use connected components labeling to segment the multiple areas of predicted pneumonia, and finally, a bounding box is simply drawn around every connected component.
Emphasizing accuracy and efficiency of image analytics, OSP utilized 'Keras,' a high-level API with Tensorflow backend which helped us to build neural networks quickly, without any intricacies. Keras uses a highly complex framework on the back-end which you get to define. The neural network we decided upon, consists of some residual blocks with convolutions and multiple downsampling blocks with max pooling to reduce its dimensionality. At the end of the created neural network system, a single upsampling layer smartly converts the output to the similar shape as the input.
A sophisticated convolutional neural network to localize pneumonia affected areas.
We used 140,000 total images after normalizing and downscaling them to a standard deviation and DICOM format in ImageNet training set.
Our advanced CNN model works on some random weights and zeroes bias, to first learn the image is of pneumonia or not to activate the further image analysis.
The chest x-rays tagged as pneumonia contain multiple opacities which are tracked by customized object localization algorithms and denoted by bounding-boxes (one per opacity).
Four parameters define a bounding box for every lung opacity. They are x-min, y-min, width, and height.
To train the model with super accuracy, we ran up to 25 Epochs and used the XML generator program to splits the data into three different datasets based on the number of bounding boxes. The bounding box distribution is analyzed to get a better idea of the visual signals for lung opacity. The processed x-ray images are passed through a Machine Learning Model. We leveraged TesnorFlow to run the model on a neural network to generate a pickle file that can be used for further predictions.
Accurate pnuemonia predictions to enhance radiologist's diagnosis performance.
Our deep learning-driven chest x-ray analytics model is a multi-layered convolutional neural network.
After inputting a chest x-ray image, it outputs the total probability of pneumonia accurately.
The output is provided in the format of a heatmap indicating the areas affected by pneumonia with bounding boxes around them.
OSP used dense connections and batch normalization to make the optimization of such a deep network easily tractable.
The image analysis model is trained to identify lung opacities with up to 60-65% accuracy. The relative sensitivity and specificity of this model is 60% and 70% respectively. The Machine Learning Model was trained with up to 1,30,000 images. The model is frequently retrained based on the new data that is received by the system to ensure reinforcement learning and enhanced accuracy. On comparison of The trained model's prediction vs. the ground truth, we observed that our model's prediction performance was slightly higher than the predictions made by the radiologists. This deep learning based pneumonia prediction solution can help specialists to make an accurate medical diagnosis.
Predicting the business future with advanced healthcare BI.
Unlocking new data patterns to transform healthcare.
Meaningful patient engagement for personalized healthcare delivery.
Powering healthcare innovation with artificial intelligence.
"Yes, I would certainly recommend their services because they were diligent and the offered price was very reasonable which is a challenge these days to get a great product at excellent pricing."
Forging alliances with top global technology companies to find a new and better way to solve complex healthcare challenges.
The experience and dedication of the people working at OSP Labs make everything we do possible. If you’re passionate about having a powerful impact on transforming healthcare, we’d love to hear from you.Talk to us