6. Performance of the global detection approach

For each indication identified by the U-Net and labeled in the 3D volume, a set of three images centered on the indication is sent to the CT-Casting-Net for classification. Only indications classified as defects three times are retained for further processing.

This approach was validated on a set of 6 tomographic volumes not used for training. The performances obtained are detailed in table 4 . The number of objects before classification corresponds to the number of indications segmented by the U-Net: this number is very high due to the over-segmentation adopted. The number of objects after classification corresponds to the number of indications classified as defects by the CT-Casting-Net. This number is much smaller, since all indications classified as false alarms are removed. The probability of detection...