Rapid image recognition of body parts scanned in computed tomography datasets

Abstract

Aim

Automatic CT dataset classification is important to efficiently create reliable database annotations, especially when large collections of scans must be analyzed.

Method

An automated segmentation and labeling algorithm was developed based on a fast patient segmentation and extraction of statistical density class features from the CT data. The method also delivers classifications of image noise level and patient size. The approach is based on image information only and uses an approximate patient contour detection and statistical features of the density distribution. These are obtained from a slice-wise analysis of the areas filled by various materials related to certain density classes and the spatial spread of each class. The resulting families of curves are subsequently classified using rules derived from knowledge about features of the human anatomy.

Results

The method was successfully applied to more than 5,000 CT datasets. Evaluation was performed via expert visual inspection of screenshots showing classification results and detected characteristic positions along the main body axis. Accuracy per body region was very satisfactory in the trunk (lung/liver >99.5% detection rate, presence of abdomen >97% or pelvis >95.8%) improvements are required for zoomed scans.

Conclusion

The method performed very reliably. A test on 1,860 CT datasets collected from an oncological trial showed that the method is feasible, efficient, and is promising as an automated tool for image post-processing.