Detection and Classification of Objects in Three-Dimensional Images Using Deep Learning Methods

The issue of automatic object detection and categorization in three-dimensional single-channel raster images is considered. The objects may have low contrast and substantial shape variability, making it challenging to explicitly construct a model. The proposed solution employs machine learning techniques based on a labeled database of use scenarios. A two-step algorithm is presented, with the first stage being the detection of objects within the image and the second being the reduction of false positives and object categorization. A deep learning approach is applied with a single input and trained for the simultaneous solution of multiple tasks. The practical goal of developing a clinically viable automatic decision support system to detect and classify rib fractures based on computed tomography (CT) scans is solved. Computational experiments are conducted on the publicly available RibFrac dataset. The proposed system is shown to achieve a detection sensitivity of 0.935, with an average number of false positive predictions per image of 4.7. The resulting algorithm is compared with the existing methods using quantitative measures.