Fig 1.
Flow diagram for patient selection.
The diagram shows the initial case selection and final distribution of study cases into the training set and test set. Jan = January, Mar = March.
Fig 2.
Flow diagram of our proposed CAD algorithms.
TP = true positive, FP = false positive, ANN = artificial neural network.
Fig 3.
Six spherical templates by sizes (2, 3, and 4 mm) and types (solid and inner-hole).
Fig 4.
Example of an ANN for FP reduction of BM candidates using computer features.
Table 1.
Clinical characteristics of the patients.
Fig 5.
Bar graph of the nodule size distributions in the training and test sets.
The relative frequency of nodules with diameters of 1 to 3 mm differed significantly between the two groups (p = 0.01).
Fig 6.
Examples of CAD results using algorithm A.
A–D: Examples of the correct detection of BM by CAD software. E–H: Examples of the incorrect detection (FPs) by CAD software. Common sources of FPs included the cortical vessel (F), dural sinus (G), and choroid plexus (H).
Table 2.
Comparison of the nodule detection performances of algorithm A and algorithm B.
Table 3.
Comparison of the reviewers’ nodule detection performances.
Fig 7.
3D gradient-echo contrast-enhanced T1-weighted MR images in an 81-year-old female patient with metastatic lung cancer.
A and B: Axial (A) and coronal (B) images show a tiny enhancing nodule at the left inferior temporal gyrus (arrowhead). This nodule was missed by all four reviewers but was successfully detected by CAD. C: On the navigation MR image for a gamma-knife surgery performed 2 days after (A) and (B), the nodule showed no interval changes. D: On the follow-up MR image taken after 3 months, the nodule disappeared.