Fig 1.
Typical findings on MRI - T1 MPRAGE: A: basal enhancement, vasculitis with thickened vascular walls, and hydrocephalus with dilated lateral ventricles; B: basal enhancement; C: tuberculoma (arrow); D: vasculitis
Table 1.
Modified MRC Grading for TBM.
Table 2.
Baseline characteristics, by HIV status.
Fig 2.
Proposed Fused CNN architecture.
The network comprises of two branches: Imaging feature extractor (1, green-top) and Non-imaging feature extractor (2, red-bottom). The output of (1) is a 512x[3x3x3] array, which was subsequently average pooled into a 512x1 vector. This vector is compressed into a 24x1 vector using a fully-connected layer before being concatenated with the 7x1 vector outputted from branch (2). The 31x1 concatenated vector is fed into a Classifier (3, violet-bottom), which performs the classification tasks. During the training, classification was performed on both branch-specific latent vectors (forming and
) and from the Classifier itself (forming
). However, in validation, only prediction from the Classifer was taken into account.
Fig 3.
Patient recruitment profile.
Table 3.
Baseline characteristics, by data spliting to 70% training/validation and 30% testing.
Table 4.
Performances of the non-imaging, imaging-only, and fused model after fine-tuning, on the validation and test set. For each type of model, the performance metrics (AUC and IPA) and their variation were estimated based on n = 10 models optimized on n = 10 folds.
Fig 4.
Performance of our model against validation and test datasets.
Each model was trained n=10 times on 10 cross-validated folds of the 2 repeated 5-fold cross-validations. A, B: Average ROC curves on validation and test datasets, stratified by models. C, D: Box plots of AUCs on validation and test datasets across folds, stratified by models. E, F: Box plots of scaled Brier scores on validation and test datasets across folds, stratified by models.
Table 5.
Performances of the non-imaging, imaging-only, and fused model on test set (N=56), stratified by HIV status. For each type of model, the performance metrics (AUC and IPA) and their variation were estimated based on n=10 models optimised on n=10 folds.
Fig 5.
Sample saliency maps showing several regions guiding the model, on sagittal and transversal planes.
The redder the pixel, the more it contributed to the prediction. The model focused on the corpus callosum (A, green arrow), cerebellum (A, purple arrow), brain stem (B, green and purple arrows), and temporal lobe around Sylvian fisure (D, green arrows). The enhanced basal meninges and tuberculomas (orange arrows) did not provide a strong contribution to the decision.