The purpose of this study is to introduce machine learning methods as a tool for automatic labeling of CT scan images according to their position in the human body. Nowadays, image databases have become very large requiring, for instance, more than 1GB for just a single full body CT scan. The acquired images are typically numbered in an increasing manner from head to foot, but not necessarily on a comparable manner as distance between two scans (i.e. 1cm, 1.5cm, etc) can differ from set to set.
A good example is the analysis of the effectiveness of some chemotherapeutic agents, normally looking for reductions of neoplastic mass. In these clinical trials, the object is a relatively small organ, and we could be working with many patients, all of them with non-labeled CT scan images. To localize one patients` pancreas among all the CT scans can be a tedious task for the physician.
Previous studies have proposed to use a two step k-NN search to solve this problem. In this work we apply an advanced Machine Learning method name Diffusion Map to tackle this task. The diffusion maps compares between two CT scans with a metric that is invariant to the features that are unique to each person, meaning that the comparison is done based on the intrinsic features that define the location in the body. Last, we compare our results with the previous methods and prove the robustness of the proposed methods.
