Accurate classification of brain tumours on magnetic resonance imaging is demanding due to variable appearances and the need for specialist review under time pressure. An integrated deep learning approach, RDXNet, was created to support consistent identification across common tumour categories. The model combines complementary convolutional architectures within a single pipeline and is trained on a large set of publicly available images. The aim is to balance expressiveness with generalisation, reduce overfitting and provide visual explanations that make outputs easier to interpret. Reported performance indicates clear gains over individual backbones and simpler combinations, with accuracy reaching a high level while maintaining attention to clinically relevant error patterns. The workflow is designed to be reproducible, with clearly defined data preparation steps and validation across multiple splits to assess robustness under different sampling conditions.
Integrated Design and Training Strategy
RDXNet integrates three established convolutional networks in a hybrid configuration to leverage their differing strengths. By fusing feature representations from these backbones into a shared head for multiclass prediction, the approach seeks to capture both fine local patterns and broader contextual cues that are useful for differentiating tumour types. Transfer learning is used so that pre-trained weights provide a strong starting point, while new layers adapt the combined features to the specific task. Regularisation techniques are applied to address overfitting, which was more apparent when single networks were trained in isolation.
The training protocol tracks not only accuracy and loss but also error characteristics across classes. This reflects the need to manage false negatives and other misclassifications that matter in clinical review. To assess stability, performance is checked across multiple data splits using K-Fold cross-validation in addition to a standard held-out set. Visual explainability is incorporated through gradient-based class activation maps that highlight image regions most influential for each prediction. These heatmaps are intended to aid scrutiny by showing whether the model focuses on tumour-relevant structures rather than background artefacts.
Cohesive Dataset and Preprocessing
The dataset aggregates three public sources and comprises 7,023 high-resolution images spanning four categories: glioma, meningioma, pituitary tumour and no tumour. Images are consistently formatted to facilitate integration and are partitioned into separate training and testing sets to allow an unbiased evaluation of performance. The class distribution covers each category in meaningful numbers to help the model learn discriminative patterns without skewing heavily toward a single label.
A structured preprocessing pipeline prepares the images for learning and supports reproducibility. Steps include removal of corrupted or near-duplicate scans, isolation of brain tissue, intensity normalisation and resizing to a common input shape. Labels are encoded for multiclass prediction and data augmentation introduces controlled variability. Two alternative paths are compared: a comprehensive preprocessing workflow and a raw-image pipeline. The processed inputs lead to clearer boundaries between classes and more stable training dynamics, whereas models trained on unprocessed images show greater instability and a tendency to overfit. This standardisation also simplifies integration of the different backbones by ensuring that each receives uniformly prepared data.
Performance, Validation and Interpretability
The hybrid configuration demonstrates stronger generalisation than individual networks and pairwise fusions. With the full preprocessing pipeline, reported accuracy reaches 94%, whereas training on raw images yields substantially lower results. Gains are also reflected across other commonly reported metrics, pointing to a more balanced error profile. Cross-validation across multiple splits indicates that the improvement is not confined to a single partition, supporting the claim of robustness under different sampling conditions.
Interpretability is addressed through Grad-CAM visualisations that reveal where the model concentrates attention when assigning a class. These maps provide a qualitative check that the network is focusing on regions consistent with tumour presence or absence rather than noise. Together with the hybrid design and disciplined data preparation, the visual explanations help frame the model as a decision support component that can be interrogated, rather than a black box. The authors also note that the backbone configuration and preprocessing framework could be adapted to related tasks such as localisation or segmentation, offering a path for extension without altering the core approach to feature fusion and validation.
A hybrid deep learning pipeline that fuses complementary backbones, standardises inputs and embeds visual explanations delivers high multiclass accuracy for brain tumour classification. By combining robust preprocessing with cross-validated training and heatmap-based interpretability, the approach improves reliability over single models while maintaining transparency suitable for clinical review. The results point to a practical direction for computer-aided diagnosis that can support timely triage and consistent image assessment within existing workflows.
Source: Healthcare Analytics
Image Credit: iStock
