Table 2 Comparison of segmentation performance for the five DL training methods for all scans in the testing set.

From: Large-scale investigation of deep learning approaches for ventilated lung segmentation using multi-nuclear hyperpolarized gas MRI

Experimental DL methods

Evaluation metrics: median (range)

DSC

Avg HD (mm)

HD95 (mm)

XOR

Train on 3He

0.961 (0.765, 0.981)

2.335 (35.91, 0.644)

10.00 (140.9, 1.934)

0.079 (0.613, 0.037)

Train on 129Xe

0.964 (0.886, 0.983)

1.341 (3.911, 0.675)

4.809 (15.90, 1.875)

0.072 (0.253, 0.035)

Train on 3He, fine-tuned on 129Xe

0.963 (0.892, 0.983)

1.384 (4.628, 0.636)

4.971 (29.80, 1.934)

0.075 (0.238, 0.034)

Train on 129Xe, fine-tuned on 3He

0.968 (0.842, 0.983)

1.483 (10.84, 0.596)

4.935 (67.85, 1.563)

0.066 (0.372, 0.034

Combined 3He and 129Xe training

0.971 (0.886, 0.983)

1.234 (5.630, 0.594)

4.193 (52.70, 1.875)

0.059 (0.255, 0.035)

  1. Medians (ranges) are given; the best result for each metric is in bold.
Back to article page