Classification of arthritis from skeletal Tc-99m MDP images with artificial neural networks

Full text: Artificial Neural Networks (ANNs) are used extensively in the pharmaceutical sciences where they excel in tasks of pattern recognition, classification and prediction. In the Nuclear Medicine literature ANNs are being developed as diagnostic assistants to classify and interpret cardiac bull's eyes, cerebral perfusion defects and lung scans for PE. This study follows on from previous work on ANN interpretation of metastatic bone disease from whole body images of patients undergoing Tc-99m MDP bone scans. The original work could classify metastatic disease, Pagets' disease and normal scans with encouraging results but not distinguish arthritic regions from metastatic disease or normal disease free images. A new program was written to extract skeletal uptake information with the emphasis on joints where arthritis is commonly seen. Mean count density and standard deviations from 17 image sites per patient (76 patients) were submitted to ANNs for classification. ANNs correctly classified 100% of all scans on training. Upon cross validation testing the ANNs correctly classified 100% of the patients with arthritis and 83% of the normal scans. These encouraging findings have identified how improvements can be made in the earlier work and caused us to pursue a more extensive data set so that future work can now combine this study with metastatic disease, Pagets' disease and normal scans. The eventual aim of this line of investigation is to provide a 'diagnostic assistant' for Nuclear Medicine physicians in training