BREAST DTS

CAD BREAST DTS stand-alone performance

Methods

Statistics

The results shown below indicate the detection performances, in terms of sensitivity and specificity on a independent testing sets, used to evaluate the system performances (training was performed on a different dataset).
The sensitivity is indicated per lesion, thus considering as correctly identified a lesion prompted by the CAD in one of the two (MLO or CC) tomosynthesis 3D projections. The specificity of the CAD is represented by the number of false positives prompt per view, i.e. the number of CAD prompts (lesion candidates) that do not correspond to a true lesion per tomosynthesis 3D volume.
The CAD performances are calculated at an operating point set on a different dataset, as a good trade-off between sensitivity and specificity. 

Datasets

All exams were acquired on a Hologic Selenia Dimensions Mammography Device, installed in the San Giovanni Battista University Hospital of Turin (Italy).
A breast imaging specialist – who was aware of the final reports – annotated the findings with a specifically designed software, by drawing a three-dimensional bounding-box around the finding. 

Results

Microcalcification Clusters Detection

The testing dataset is made up of 119 patients (31 lesions), for a total of 433 tomosynthesis volumes.

     Sensitivity per lesion: 94% (C.I. 95%: 78-99%)
     Average number of false positives per view: 1 (std. dev. : +/- 2)
     Median of false positive per view: 0 (first and third quartile: [0,1])

Masses Detection

The testing dataset is made up of 82 patients (39 lesions), for a total of 290 tomosynthesis volumes.

     Sensitivity per lesion: 87% (C.I. 95%: 73-94%)
     Average number of false positives per view: 5 (
std. dev. : +/- 3)