Basic (Target Identification)
Accurate automated organ and disease feature segmentation is a challenge for medical imaging analysis. The pancreas, for example, is a small, soft, organ with low uniformity of shape and volume between patients. Because of the lack of uniform image patterns, there are few features that can be used to aid in automated identification of anatomy and boundaries. Segmentation of high variability features is uniquely difficult for a computer to perform. Due to these difficulties, high variability anatomical features are currently analyzed and determined only by trained physicians who can read the images. Another challenge is that there is a shortage of trained physicians relative to the amount of image data generated. While computer automation may help solve many limitations for human image analysis, which is time consuming and labor intensive, it has been difficult to achieve.
To help solve some of these challenges, researchers at the National Institutes of Health Clinical Center (NIHCC) have developed a technology that trains a computer to read and segment certain highly variable images features, such as the pancreas. This analysis is done by employing Holistically-Nested Convolutional Neural Network (HNNs) and deep learning. The resulting biomarkers are far more precise compared to other approaches and outperform current methods for automated image localization and segmentation of high variability image features. The training methods may be generalizable to enable automation of segmentation for many high variability image structures, such as tumors and diseased organs. This advancement has application for improving computer assisted diagnostic capabilities, and disease monitoring and surgical planning abilities for many diseases.