Glioblastoma is one of the most aggressive brain tumours, and analysing its histopathology images is essential for accurate diagnosis and prognosis. However, identifying distinct tumour structures in stained tissue sections remains a challenging and time-consuming task for pathologists. In this paper, we present a deep learning approach for multi-class classification of nine distinct tumour sub-regions in H&E-stained histology slides, developed in the context of the BraTS-Path 2025 challenge. We leverage transfer learning with MobileNetV2 as a baseline, then progressively improve it through advanced optimisation techniques. We further validated the model’s generalisability using rigorous 5-fold cross-validation. The experimental results demonstrate a substantial improvement over the baseline: the optimised pipeline achieves high overall accuracy (98.54%) and robust class-wise performance (macro-averaged F1-score 95.05%). The optimised model achieves robust and consistent outcomes for each of the 9 tumour classes, as evidenced by per-class receiver operating characteristic curves and confusion matrices.