Fig. 1: Qualitative comparison of segmentation (top) and detection (bottom).
From: GLANCE: continuous global-local exchange with consensus fusion for robust nodule segmentation
For each case, we show the original image, ground truth, ours, and a series of baselines arranged left-to-right in progressively worse visual quality relative to (c)(k)ours. Detection baselines: (d)LN-DETR, (e)NDLA, (f)AWEU-Net, (g)CSE-GAN, (h)EHO-Deep CNN, segmentation baselines: (l)UnetTransCNN, (m)BRAU-Net++, (n)CT-UFormer, (o)UNETR++, (p)Improved V-Net; Our method yields sharper boundaries and fewer false positives/negatives, with competing methods exhibiting increasing boundary erosion, missed lesions, and spurious responses to the right.
