A Comparative Study of a Hybrid CNN–Transformer Framework for Multimodal MRI-Based Brain Tumour Segmentation

Brain tumour segmentation in MRIs is a difficult task due to varying appearance of tumours between sequences, different regions in space and among cases. The CNNs are useful for efficient local feature extraction and decoding in multiple scales, whereas Transformer attention provides the ability to capture the relations between features in a wider sense. This work explores a compact U-Net architecture versus the hybrid CNN–Transformer model for a notebook on the basis of four modalities from the BraTS-GLI 2023 training collection: T1n, T1c, T2w and T2-FLAIR (T2f). Four-class pixel-wise segmentation of 128 × 128 resolution was performed by both models. Training was conducted using AdamW optimiser with weighted cross-entropy and foreground soft-Dice loss functions. Validation macro foreground Dice was used as a metric for checkpoints selection. Testing was conducted on 194 out-of-bag case IDs and six randomly chosen slices per case (total number of slices = 1,164 slices in total); therefore, the experiment is a selected-slice 2D evaluation rather than full-volume 3D segmentation. The two models scored a macro foreground Dice score of 0.7759 and 0.8132 on the chosen test slices, respectively, while their pooled whole-tumor Dice score was 0.8585 and 0.9139. The pixel accuracy scores of the U-Net and the hybrid were 0.9953 and 0.9969, respectively. A paired case-level comparison using 10,000 bootstrap samples, with the Wilcoxon signed-rank test and Holm correction showed significant differences that favoured the hybrid in dice scores of encoded label 2, macro foreground Dice, whole-tumor Dice, and pixel accuracy. However, there were no statistically significant differences between the two models for encoded labels 1 and 3 after Holm correction. The hybrid model contained 361,204 trainable parameters while those of the U-Net were 1,942,772. Although these results are promising, caution needs to be exercised since repeated numerically labelled data instances require identity verification, label semantics have to be validated against the official dataset documentation, no external dataset was used, and there is poor overlap of Grad-CAM with the reference foreground.