Computer Aided Detection of Lytic Bone Metastases in the Spine using Routine CT Images

Bone metastases occur when cancer cells from the primary tumors relocate to the bones. This paper presents a computer aided detection system to identify lytic bone metastases in the thoracolumbar spine using routine chest/abdominal CT scans. The spine is segmented using a hybrid technique of thresholding, mathematical morphology, directed graph search, and a four-part vertebral model. A watershed algorithm is applied to search for lytic detections that match size, shape, location and intensity criteria. To reduce the number of false positives, we computed 26 quantitative features (density, shape and location) and fed them to a support vector machine (SVM). We have tested our method on 50 data sets and 90 bone metastases. The per-patient sensitivities are 85.7%, 91.7% and 100% for bone lesions>0 cm², >0.2 cm ² and >0.8 cm² in the test sets, with 5.7 false positive per-patient