Table 5.

Universal lesion detection (ULD) task results and inference time on the selected small DeepLesion dataset. Sensitivity (%) at various FPs per image and the mean average precision (mAP) are used as the evaluation metrics. $std$ denotes the standard deviation value. $IT$ denotes the inference time.

Evaluation Metric	Sen@0.5 ± std	Sen@1 ± std	Sen@2 ± std	Sen@4 ± std	Sen@8 ± std	Sen@16 ± std	mAP ± std	IT(ms) ± std
Faster R-CNN [25]	58.93 ± 1.87	68.46 ± 0.18	76.69 ± 0.99	82.27 ± 1.75	85.98 ± 2.02	88.52 ± 1.88	52.80 ± 1.19	200 ± 3.79
Original R-FCN [21]	57.89 ± 0.64	68.69 ± 0.56	76.60 ± 1.17	82.12 ± 1.38	86.28 ± 1.73	88.61 ± 1.79	50.13 ± 1.02	222 ± 31.77
3DCE, 9 slices [1]	62.70 ± 0.71	72.20 ± 0.46	79.85 ± 1.58	84.51 ± 1.76	87.43 ± 1.73	89.67 ± 1.85	57.09 ± 1.93	239 ± 24.83
Dense DAL 3DCE R-FCN, 9 slices [3]	63.32 ± 1.10	72.79 ± 0.95	80.94 ± 2.29	85.93 ± 2.29	88.88 ± 2.50	90.82 ± 2.38	58.28 ± 0.77	235 ± 3.51
DA-SHT Dense 3DCE R-FCN, 9 slices (ours)	64.86 ± 1.43	74.39 ± 1.01	81.41 ± 2.08	86.04 ± 2.38	89.26 ± 1.82	91.38 ± 1.77	59.22 ± 1.19	241 ± 2.09