Figure 1.

(A) Schematic of the scan-to-prediction pipeline for molecular subtype classification. The pipeline inputs the raw T2W MRI scan and outputs the mutation class prediction. (B) Input and output depiction of the segmentation model from stage 1 of the pipeline. The segmentation block also involves registration and preprocessing of the input scan. The output consists of the preprocessed input MRI scan along with the co-registered segmentation mask. (C) Flow diagram of the TransferX training block and approach. The TransferX algorithm is employed to train three individual subtype classifiers (BRAF V600E, Wild-type, BRAF Fusion). (D) The model architecture of individual binary molecular subtype classifier. (E) Schematic of consensus decision block. The block inputs the classification outputs and corresponding scores from the three individual subtype classifiers and fits them into a consensus logic, and outputs the final predictions. The mutational class predictions are output sequentially where the input is first checked for wild-type or non-BRAF class first. If the input doesn’t belong to wildtype or non-BRAF class, then the logic progresses to check the BRAF mutation class with BRAF Fusion checked first then followed by BRAF V600E. T2W: T2-weighted.