Information Technology Reference
In-Depth Information
11.2 Sclerotic Metastasis Detection Results
Figure
11
shows one example of sclerotic metastasis CAD results. It shows the
output of each stage in the CAD system. A total of 552 sclerotic lesions were
electronically location marked and segmented in the ground truth data set of 50
cases, for an average of 11.2 lesions per patient [standard deviation (SD): 10.8], and
a range of 0
59 lesions per patient (Table
3
). The average number of lesions per
patient detected by the CAD system on this data set was 9.9 (SD: 10.1) with a range
0
-
57 lesions detected per patient. All 50 case CT studies accumulated during the
routine clinical review phase of the study appeared positive for lesions upon
qualitative visual assessment.
Figure
12
shows the FROC analysis for the sclerotic metastasis CAD system.
The operating point value for the training set was chosen to target a sensitivity of
90 % or greater, while maintaining a clinically reasonable false positive rate. This
same operating point was then used for the testing set. The CAD system sensitivity
for the training set was 0.90 (95 % con
-
dence interval: 0.86, 0.92) at a false
positive rate (FPR) of 8.9 (95 % CI: 6.3, 12.0) per patient. The testing set sensitivity
was 0.84 (95 % CI: 0.80, 0.87) with an FPR of 11.3 (95 % CI: 9.3, 13.6) per patient.
Fig. 11 Sclerotic metastasis CAD. 75 year-old patient with prostate cancer. a 3-D segmented
spine with ground truth lesions marked in light blue. b Detections (red) from 3D merging
algorithm, prior to detection filter screening (red). c Red candidate lesions remaining after
screening by detection filter. Blue detections eliminated by filter. d, e Lesions remaining post SVM
classification with cutoff 0.48 (d) and 0.55 (e). Green TP lesions. Red FP lesions. False positive
detections due to degenerative change and partial volume vertebral end plates
