Article Figures & Data
Figures
- FIG 1.
Comparison of model-with-radiologist performance. A, ROC curve for scan-level tumor detection. Model, individual radiologist, and average radiologist performance are indicated with crosshairs. B, Model and average radiologist performance for tumor subtype classification results. Error bars represent standard error among radiologists.
- FIG 2.
CAMs depicting the areas of the input slice that the model preferentially emphasizes when predicting tumor subtype on individual scan slices. The upper row of each subpanel shows the T2 slice with tumor areas manually denoted (upper left) and CAM overlay of the most confident prediction of the model (upper right). The lower row of each panel shows less confident predictions. Examples of correct predictions of PA (A) and MB (B) and incorrect predictions of PA (C) and MB (D) are shown.
- FIG 3.
Learned feature vectors were reduced to 2D and visualized using principal component analysis (PCA) (A) and t-SNE (B). DMG has the most distinctive feature space, followed by PA and MB. EP has the least distinctive feature space and overlaps with MB.
Tables
Institution 1 Institution 2 Institution 3 Institution 4 Institution 5 MB 90 117 20 30 41 DMG 85 0 0 45 21 EP 42 41 41 22 8 PA 129 0 0 45 26 ↵a One hundred eighty-six patients with no T2 sequences or only postintervention imaging were excluded.
- Table 2:
A total of 739 scans were distributed into a training set, a validation set, and a held-out test set
Training Validation Test Total MB 242 34 55 331 DMG 88 10 24 122 EP 83 13 15 111 PA 114 20 41 175 Total 527 77 135 739 - Table 3:
Comparison of tumor detection and classification results between the deep learning model and radiologistsa
Tumor Detection Tumor Classification Sensitivity Specificity P Accuracy F1 Score P Model 0.96 1.00 – 0.92 0.80 – Radiologist average 0.99 0.98 – 0.87 0.75 – Radiologist A 1.00 1.00 .06 0.95 0.89 .09 Radiologist B 0.99 0.97 1.00 0.89 0.79 .24 Radiologist C 0.99 0.98 1.00 0.79 0.61 <.01 Radiologist D 0.98 0.98 .73 0.84 0.70 <.01 Note:— –indicates n/a.
↵a P value calculated using the McNemar test comparing the model with individual radiologists.
Jump to section
Related Articles
Cited By...
- Pediatric brain tumor classification using deep learning on MR-images with age fusion
- Deep Learning Outperforms Classical Machine Learning Methods in Pediatric Brain Tumor Classification through Mass Spectra
- Pediatric brain tumor classification using deep learning on MR-images from the childrens brain tumor network
- Pediatric brain tumor classification using deep learning on MR-images from the childrens brain tumor network
- Pediatric brain tumor type classification using deep learning on MR images from the childrens brain tumor network
- Radiomics Can Distinguish Pediatric Supratentorial Embryonal Tumors, High-Grade Gliomas, and Ependymomas