Introduction
Material and methods
Data collection
Radiomic feature extraction and image processing
Datasets construction and vViT setting
Statistical analysis
Results
Parameter | Training dataset 122 patients, 1,570 images | Test dataset 30 patients, 484 images |
---|---|---|
Mean age (95%CI), years | 61.4 (37.5–85.5) | 62.1 (39.3–85.0) |
Sex, n (%) | ||
Men | 86 (70.5) | 27 (90.0) |
Women | 36 (29.5) | 3 (10.0) |
Pathologic diagnosisa, n (%) | ||
Glioblastoma, IDH-wildtype | 118 (96.7) | 30 (100.0) |
Astrocytoma, IDH-mutant | 4 (3.3) | 0 (0.0) |
WHO gradea, n (%) | ||
4 | 122 (100.0) | 30 (100.0) |
MGMT promoter methylation, n (%) | ||
Methylated | 88 (72.1) | 20 (66.7) |
Unmethylated | 34 (28.9) | 10 (33.3) |
Images per patient, n (95%CI) | 12.86 (11.4–14.3) | 16.1 (13.4–18.9) |
Image-based analysis
Statistic (95%CI) | Total model output | Class token sector | Demographic sector | CE-T1WI ET radiomic sector | CE-T1WI NCR radiomic sector | CE-T1WI ED radiomic sector | T2WI ET radiomic sector | T2WI NCR radiomic sector | T2WI ED radiomic sector | CE-T1WI sector | T2WI sector |
---|---|---|---|---|---|---|---|---|---|---|---|
(a) Image-based analysis for the test dataset | |||||||||||
Accuracy | 0.764 (0.743–0.782) | 0.736 (0.714–0.754) | 0.758 (0.737–0.776) | 0.767 (0.745–0.784) | 0.762 (0.741–0.780) | 0.767 (0.745–0.784) | 0.760 (0.739–0.778) | 0.762 (0.741–0.780) | 0.758 (0.737–0.776) | 0.758 (0.737–0.776) | 0.731 (0.710–0.750) |
Sensitivity | 0.612 (0.579–0.641) | 0.566 (0.533–0.597) | 0.607 (0.574–0.637) | 0.603 (0.570–0.633) | 0.616 (0.583–0.645) | 0.607 (0.574–0.637) | 0.624 (0.591–0.653) | 0.616 (0.583–0.645) | 0.607 (0.574–0.637) | 0.620 (0.587–0.649) | 0.492 (0.460–0.524) |
Specificity | 0.917 (0.892–0.930) | 0.905 (0.879–0.918) | 0.909 (0.883–0.922) | 0.930 (0.905–0.941) | 0.909 (0.883–0.922) | 0.926 (0.901–0.937) | 0.897 (0.870–0.911) | 0.909 (0.883–0.922) | 0.909 (0.883–0.922) | 0.897 (0.870–0.911) | 0.971 (0.951–0.977) |
PPV | 0.881 (0.846–0.899) | 0.856 (0.819–0.877) | 0.870 (0.834–0.889) | 0.896 (0.861–0.912) | 0.871 (0.836–0.890) | 0.891 (0.856–0.908) | 0.858 (0.823–0.878) | 0.871 (0.836–0.890) | 0.870 (0.834–0.889) | 0.857 (0.822–0.877) | 0.944 (0.907–0.956) |
NPV | 0.703 (0.674–0.726) | 0.676 (0.648–0.700) | 0.698 (0.670–0.722) | 0.701 (0.673–0.724) | 0.703 (0.675–0.726) | 0.702 (0.674–0.725) | 0.705 (0.676–0.728) | 0.703 (0.675–0.726) | 0.698 (0.670–0.722) | 0.702 (0.674–0.726) | 0.656 (0.630–0.680) |
F-score | 0.722 (0.682–0.762) | 0.682 (0.640–0.723) | 0.715 (0.675–0.756) | 0.721 (0.681–0.761) | 0.722 (0.682–0.761) | 0.722 (0.682–0.762) | 0.722 (0.683–0.762) | 0.722 (0.682–0.761) | 0.715 (0.675–0.756) | 0.719 (0.679–0.759) | 0.647 (0.604–0.689) |
AUC-ROC | 0.828 (0.792–0.862) | 0.797 (0.758–0.835) | 0.787 (0.747–0.829) | 0.783 (0.743–0.825) | 0.793 (0.753–0.835) | 0.790 (0.749–0.832) | 0.785 (0.744–0.828) | 0.803 (0.763–0.842) | 0.800 (0.760–0.840) | 0.830 (0.794–0.863) | 0.813 (0.777–0.852) |
Logarithmic loss | 1.34 (1.34–1.35) | 2.07 (2.06–2.07) | 2.39 (2.39–2.39) | 2.62 (2.61–2.62) | 2.53 (2.52–2.53) | 2.50 (2.50–2.50) | 2.49 (2.49–2.50) | 2.50 (2.50–2.50) | 2.58 (2.57–2.58) | 0.955 (0.952–0.958) | 1.59 (1.59–1.60) |
Cohen’s Kappa score | 0.529 (0.510–0.548) | 0.471 (0.451–0.491) | 0.517 (0.497–0.536) | 0.533 (0.514–0.552) | 0.525 (0.506–0.544) | 0.533 (0.514–0.552) | 0.521 (0.502–0.540) | 0.525 (0.506–0.544) | 0.517 (0.497–0.536) | 0.517 (0.497–0.536) | 0.463 (0.443–0.483) |
(b) Patient-based analysis for the test dataset | |||||||||||
Accuracy | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.833 (0.714–0.877) | 0.867 (0.747–0.903) |
Sensitivity | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) | 0.600 (0.395–0.749) |
Specificity | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 0.950 (0.788–0.967) | 1.00 (0.839–1.000) |
PPV | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 0.857 (0.535–0.927) | 1.00 (0.610–1.000) |
NPV | 0.826 (0.682–0.877) | 0.826 (0.682–0.877) | 0.826 (0.682–0.877) | 0.826 (0.682–0.877) | 0.826 (0.682–0.877) | 0.82E-T1TW (0.682–0.877) | 0.826 (0.682–0.877) | 0.826 (0.682–0.877) | 0.826 (0.682–0.877) | 0.826 (0.682–0.877) | 0.833 (0.693–0.882) |
F-score | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.706 (0.543–0.869) | 0.750 (0.595–0.905) |
AUC-ROC | 0.840 (0.650–0.995) | 0.825 (0.636–0.992) | 0.820 (0.619–0.983) | 0.810 (0.591–0.995) | 0.825 (0.620–0.995) | 0.830 (0.620–1.000) | 0.845 (0.659–1.000) | 0.840 (0.656–0.990) | 0.815 (0.609–0.985) | 0.840 (0.656–0.995) | 0.835 (0.652–0.990) |
Logarithmic loss | 0.613 (0.561–0.665) | 0.807 (0.750–0.863) | 0.888 (0.833–0.943) | 1.05 (0.990–1.11) | 0.977 (0.920–1.03) | 1.01 (0.954–1.07) | 0.889 (0.835–0.944) | 0.925 (0.868–0.981) | 1.02 (0.959–1.08) | 0.474 (0.418–0.530) | 0.741 (0.683–0.798) |
Cohen’s Kappa score | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.595 (0.540–0.649) | 0.667 (0.618–0.715) |
Statistics (95%CI) | CE-T1WI | T2WI | ||||||||
---|---|---|---|---|---|---|---|---|---|---|
vViT | AlexNet | GoogleNet | VGG16 | ResNet | vViT | AlexNet | GoogleNet | VGG16 | ResNet | |
(a) Image-based analysis | ||||||||||
Accuracy | 0.758 (0.737–0.776) | 0.504 (0.482–0.527) | 0.494 (0.471–0.516) | 0.467 (0.445–0.490) | 0.519 (0.496–0.541) | 0.833 (0.714–0.877) | 0.531 (0.508–0.553) | 0.506 (0.484–0.529) | 0.496 (0.473–0.518) | 0.529 (0.506–0.551) |
Sensitivity | 0.620 (0.587–0.649) | 0.091 (0.078–0.117) | 0.430 (0.399–0.463) | 0.636 (0.603–0.665) | 0.698 (0.666–0.725) | 0.600 (0.395–0.749) | 0.459 (0.427–0.491) | 0.331 (0.303–0.363) | 0.888 (0.861–0.903) | 0.698 (0.666–0.725) |
Specificity | 0.897 (0.870–0.911) | 0.917 (0.892–0.930) | 0.558 (0.525–0.589) | 0.298 (0.271–0.330) | 0.339 (0.311–0.371) | 0.950 (0.788–0.967) | 0.603 (0.570–0.633) | 0.682 (0.649–0.709) | 0.103 (0.089–0.130) | 0.360 (0.331–0.392) |
PPV | 0.857 (0.822–0.877) | 0.524 (0.441–0.603) | 0.493 (0.459–0.527) | 0.475 (0.448–0.503) | 0.514 (0.486–0.541) | 0.857 (0.535–0.927) | 0.536 (0.501–0.570) | 0.510 (0.469–0.549) | 0.498 (0.474–0.521) | 0.522 (0.494–0.549) |
NPV | 0.702 (0.674–0.726) | 0.502 (0.479–0.526) | 0.495 (0.465–0.525) | 0.450 (0.412–0.491) | 0.529 (0.488–0.569) | 0.826 (0.682–0.877) | 0.527 (0.497–0.556) | 0.505 (0.477–0.532) | 0.481 (0.410–0.554) | 0.544 (0.503–0.582) |
F-score | 0.719 (0.679–0.759) | 0.155 (0.123–0.187) | 0.459 (0.415–0.504) | 0.544 (0.500–0.589) | 0.592 (0.548–0.636) | 0.706 (0.543–0.869) | 0.494 (0.450–0.539) | 0.401 (0.357–0.445) | 0.638 (0.595–0.681) | 0.597 (0.553–0.641) |
p-value | - | <0.0001 | 0.027 | <0.0001 | <0.0001 | - | <0.0001 | 0.031 | <0.0001 | <0.0001 |
AUC-ROC | 0.830 (0.794–0.863) | 0.485 (0.437-0.534) | 0.483 (0.434-0.532) | 0.504 (0.455-0.553) | 0.520 (0.471-0.569) | 0.813 (0.777–0.852) | 0.502 (0.453-0.551) | 0.535 (0.486-0.584) | 0.547 (0.498-0.596) | 0.512 (0.463-0.562) |
p-value | - | <0.0001 | <0.0001 | <0.0001 | <0.0001 | - | <0.0001 | <0.0001 | <0.0001 | <0.0001 |
(b) Patient-based analysis | ||||||||||
Accuracy | 0.833 (0.714–0.877) | 0.667 (0.554–0.742) | 0.467 (0.373–0.568) | 0.367 (0.287–0.477) | 0.400 (0.315–0.508) | 0.867 (0.747–0.903) | 0.633 (0.523–0.713) | 0.600 (0.492–0.685) | 0.333 (0.258–0.446) | 0.467 (0.373–0.568) |
Sensitivity | 0.600 (0.395–0.749) | 0.000 (0.000–0.278) | 0.000 (0.000–0.278) | 1.000 (0.722–1.000) | 1.000 (0.722–1.000) | 0.600 (0.395–0.749) | 0.300 (0.183–0.528) | 0.000 (0.000–0.278) | 1.000 (0.722–1.000) | 1.000 (0.722–1.000) |
Specificity | 0.950 (0.788–0.967) | 1.000 (0.839–1.000) | 0.700 (0.551–0.784) | 0.050 (0.033–0.212) | 0.100 (0.067–0.262) | 1.00 (0.839–1.000) | 0.800 (0.643–0.861) | 0.900 (0.738–0.933) | 0.000 (0.000–0.161) | 0.200 (0.139–0.357) |
PPV | 0.857 (0.535–0.927) | - | 0.000 (0.000–0.390) | 0.345 (0.267–0.459) | 0.357 (0.276–0.473) | 1.00 (0.610–1.000) | 0.429 (0.241–0.667) | 0.000 (0.000–0.658) | 0.333 (0.258–0.446) | 0.385 (0.296–0.503) |
NPV | 0.826 (0.682–0.877) | 0.667 (0.554–0.742) | 0.583 (0.462–0.681) | 1.000 (0.207–1.000) | 1.000 (0.342–1.000) | 0.833 (0.693–0.882) | 0.696 (0.560–0.775) | 0.643 (0.527–0.724) | - | 1.000 (0.510–1.000) |
F-score | 0.706 (0.543–0.869) | - | - | 0.513 (0.334–0.692) | 0.526 (0.348–0.705) | 0.750 (0.595–0.905) | 0.353 (0.182–0.524) | - | 0.500 (0.321–0.679) | 0.556 (0.378–0.733) |
p-value | - | 0.016 | 0.21 | <0.0001 | <0.0001 | - | 1.0 | 1.0 | <0.0001 | <0.0001 |
AUC-ROC | 0.840 (0.656–0.995) | 0.490 (0.238-0.742) | 0.530 (0.319-0.741) | 0.635 (0.424-0.846) | 0.615 (0.399-0.831) | 0.835 (0.652–0.990) | 0.535 (0.321-0.749) | 0.525 (0.317-0.733) | 0.67 (0.476-0.864) | 0.525 (0.313-0.737) |
p-value | - | 0.092 | 0.048 | 0.367 | 0.26 | - | 0.032 | 0.025 | 0.28 | 0.045 |