Almost always, our analysis consists of three stages: Segmentation and Fine-Tuning, Feature Extraction, and Presentation of Results.
Segmentation and Fine-Tuning
Given an image of some cells, we generally want to figure out where individual cells are right down to the pixel. Doing this by hand for large images with thousands of cells is inefficient as well as infeasible. This is where the deep learning comes in. A convolutional neural network is capable of rapidly producing cell predictions for very large images. Such a method requires training. Thankfully, we have a handful of pre-trained models, out-of-the-box CNNs that have been trained on specific data types. In many cases, one of these models will produce accurate results. If not, we can use a process called fine-tuning to retrain an existing model on a new set of images to perform better. We can combine one of our pre-trained models with a small amount of labeled images that it has not seen before to greatly improve prediction performance on a new dataset, yielding accurate predictions on a completely new dataset.
This process is covered more in depth in the Cell Detection, Cell Segmentation, and Cell Classification Sections.
One of the methods frequently needed in digital image analysis is Feature Extraction. Feature Extraction begins with an initial raw data set, such as a set of whole-slide images, and condenses them into a list of informative measurements, or features. Many different types of tissue samples have visual attributes that have been determined by clinicians to be of value in disease grading and diagnosis. Most of these features describe shape, structure, appearance, and distribution of cells and tissues within a sample and have been useful in histopathology imagery analysis. Computer Vision is able to provide real-world measurements of these phenomena and has rapidly been found to be an excellent tool in digital pathology analysis and Feature Extraction. Boundaries, size, length, and shape of cells can be readily identified through this process.
More on Feature Extraction can be read at our Feature Extraction Section.
Presentation of Results
Once we have gathered the data you require for your project, we can present it to you in whichever manner you prefer. Our most common request is through our Analysis Portal, an in house web application that can be used on any device connected to the internet.
If you are interested in viewing our past successes with our work in medical image analysis, navigate to our Success Stories page.