Applications of Artificial Intelligence: Medical Imaging
Medical imaging in itself is a technological wonder. The resulting images provide doctors with a way to diagnose and quickly begin treatment without having to perform invasive, exploratory procedures. The ability to take pictures inside the human body was the result of years of scientific and engineering research and development. Today, we are advancing that technology by applying artificial intelligence (AI) to pinpoint abnormalities quicker. AI can correlate data from a wide variety of images to lead doctors to new discoveries of trends and predictive markers for disease.
Dovel has developed a number of AI solutions for medical imaging that are detailed below and are available to view on our digital platform, Discover.
There are many different types of abnormalities that chest x-rays detect so doctors need ways to localize findings on x-ray images for more meaningful diagnostic results. Dovel participated in a global machine learning (ML) competition designed to find more efficient ways for studying these complex images. The event, hosted by Kaggle (a subsidiary of Google) and VinBigData, presented a challenge of localizing and classifying abnormalities in chest x-rays.
The Dovel solution approached the problem of localizing findings in four steps.
- Step 1: Train Object Detection Model for abnormalities localization.
- Step 2: Train Binary Classification Model to classify if the x-ray is normal or abnormal.
- Step 3: Filter out chest x-rays that are predicted as normal from classification model.
- Step 4: Detect and localize on abnormal x-rays filtered from Step 3.
Around 15,000 annotated chest x-rays for Chest Abnormality Object Detection were provided for training of the ML models. Dovel’s Innovation and Technology Group (ITG) data science team also used an additional 12,000+ annotated x-rays from the National Institutes of Health (NIH) to aid in the training of our ML submission.
The ITG’s data science team was awarded a bronze competition medal and placed 80 out of 1277 teams with their solution. Check out the solution here.
X-ray Segmentation: Pneumothorax Segmentation
A more specific application of AI for chest x-rays is our work on a model that identifies pneumothorax disease (a collapsed lung) in chest x-rays by segmenting the region where the pneumothorax is located. We used a deep learning semantic segmentation model called UNet++ with an EfficientNet backbone for the pneumothorax segmentation prediction. An x-ray image goes into the model and outputs the predicted pixels of the image where the pneumothorax occurs.
Microscopic Imaging: Metastatic Cancer Detection
Metastatic cancer is a cancer that has spread from the part of the body where it started (the primary site) to other parts of the body. Dovel used a deep learning model trained on around 200K annotated images to achieve 97 percent accuracy in metastatic breast cancer detection.
Data augmentation was used during the training of the model to give the model more samples to learn from and help the model be more robust to filter out other unrelated factors. Randomly horizontally flipping the image gives the model another perspective to learn from, with the same class label. Injecting random brightness and contrast to the image helps the model be more robust to brightness and contrast noise. A deep learning model was also used on the images to produce a binary prediction probability.
CT Scans: Head Hemorrhage Detection
Intracranial hemorrhage refers to bleeding within the skull that occurs when a blood vessel within the skull is ruptured or leaks. Intracranial hemorrhages account for around 10% of strokes in the U.S. Stroke is the fifth leading cause of death in the U.S. Diagnosis requires an urgent procedure, highly trained specialists, and time.
Medical Imaging analysis helps the process of identifying organs or lesions from CT scans or MRI images and can deliver essential information about the shapes and volumes of these organs. Earlier, the process of automating this procedure was done using edge detection filters and mathematical methods. However, with the advent of AI, deep learning has become the preferred capability in image processing tasks. Deep learning is indispensable to the medical industry today. However, what it has achieved is just the tip of the iceberg. Deep learning has the potential to automate significant areas of the healthcare industry.
Our solution uses two advanced deep learning models, both achieve over 98 percent accuracy. The first model is a simpler version that is used for rapid prototyping and experimentation iterations. The second model is an advanced CNN-RNN deep learning model for production deployment. CNNs can help enable highly representative, data-driven, layered hierarchical image features to be learned with sufficient training data. The main strategy that we employ when using CNNs for medical image classification is to train the CNN from scratch, using pre-trained CNN features off–the–shelf and undertaking unattended pre-training with supervised fine-tuning.
We’re continually developing and adding to Discover, be sure to check it out for additional medical imaging solutions as well as other AI, ML, and natural language processing (NLP) solutions.