Charisma: An Integrated Approach to Automatic H&E-stained Skeletal Muscle Cell Segmentation Using Supervised Learning and Novel Robust Clump Splitting Techniques

Histological images show the effects of disease and treatment on the cellular level, making them an invaluable tool for medical researchers studying these effects. Evaluating these by hand, however, is time-consuming and subjective, and, for the domain of histological muscle images, no satisfactory fully automatic solution exists. Available packages’ results often lack correct borders, causing them to identify clumped cell groups as individual cells and oversegment cells into multiple parts. The algorithms also fail to filter out non-cell tissue. Furthermore, these tools require a set of parameters to be tuned per image, severely limiting support for processing large batches of images, especially if taken under variable lighting conditions. In this research we combine various image processing and machine learning techniques to provide an image segmentation algorithm that outperforms the state of the art tools by a large margin. We first create a rough segmentation of the image using an intelligent and image-aware color thresholding to separate the cells from the background and other objects. This provides us with some individual cell segments, but may also contain unfiltered remnants of irrelevant tissue or segments that contain so-called clumps of proximate cells. In order to identify segments as belonging to one of these three categories, a discriminating SVM model is trained using a database of labeled example segments of all categories. The model is based on a number of key differentiating features, mostly based on size, shape and color of the segment. We were able to train a successful model, allowing us to accurately label each segment resulting from the initial step as clump, irrelevant, or cell.status: publishe