Asymmetric 3D Context Fusion for Universal Lesion Detection

SATr: Slice Attention with Transformer for Universal Lesion Detection

“…As shown in Table 2, our method brings promising detection performance improvements for all baselines with full training dataset. The improvements of Faster R-CNN [39], 3DCE, 3DCE w/ cBM and MVP-Net are more pronounced than those of AlignShift [9] and A3D [16]. This is because AlignShift and A3D introduce channel-fusion mechanism among different slices in backbone, thus the v value enhancement design in SATr brings less advances.…”

“…Five state-of-the-art ULD approaches [6,7,9,16,17] and one natural image [39] detection method are compared to evaluate SATr's effectiveness.…”

“…Universal Lesion Detection (ULD) in computed tomography (CT) [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18], aiming to localize different types of lesions instead of identifying lesion types [19][20][21][22][23][24][25][26][27][28], plays an essential role in computer-aided diagnosis (CAD) [29,30]. ULD is a challenging task because different lesions have diverse shapes and sizes, easily leading to false positive and false negative detections.…”

“…Mainly inspired by the clinical fact that radiologists need several adjacent slices for locating and diagnosing lesions on one CT slice, most existing ULD methods take several adjacent 2D CT slices as the inputs to a 2D network architecture [3, 4, 6-10, 12, 15-18] or directly adopt 3D network designs [10] that take 3D volume as input to extract more 3D context information. While both 2D and 3D methods have yielded great A3D [16] A3D+SATr Input CT slices cBM [17] cBM+ SATr ULD performances, the multi-slice-input based 2D detection methods are much more popular than pure 3D fashion because 2D networks benefit from robust 2D models pretrained from large-scale data whereas publicly available 3D medical datasets are not large enough for robust 3D pretraining. While achieving success in ULD, the multi-slice-input based 2D approaches have inherent limitations: (i) Weak global context modeling within each slice.…”

“…Such convolution-based fusion methods are good at dealing with local features, but they unfortunately deteriorate in handling global features among different slices. To tackle this, some ULD approaches [9,16,17] propose to reshuffle feature channels among different slices which relieves this issue to some degree. But their ability in capturing rich global representations are still weak due to the use of pure convolutional operations.…”