Deep Learning for Accurate Segmentation of Venous Thrombus from Black-Blood Magnetic Resonance Images: A Multicenter Study

Figure 3

Representative thrombus segmentation results of different networks for a patient. (a) Ground truth, (b) 3D U-Net, (c) V-Net, (d) 3D nnU-Net, (e) Cascade nnU-Net, (f) our proposed network. The first two rows show the segmentation result from the coronal plane, and the last row is to observe the full result directly from the maximum intensity projection (MIP). Red areas in the different models indicate the ground truth and the segmentation result. Yellow boxes highlight some oversegmentation errors. Blue boxes indicate the loss area of DVT segmentation results.

(a)

(b)

(c)

(d)

(e)

(f)