Preprocessing-driven early-stage brain tumour detection using baseline CNN models on clinical MRI data
Background Brain tumours represent a clinically significant category of central nervous system neoplasms, wherein early-stage detection critically determines therapeutic response and patient prognosis. Although magnetic resonance imaging (MRI) remains the gold-standard non-invasive modality for brain tumour visualization, raw MRI acquisitions are inherently degraded by multiple artefact sources including stochastic noise, intensity inhomogeneity, skull interference, and inter-scanner variability that collectively impair the discriminative capacity of computer-aided classification systems. Methods We propose a systematic, reproducible preprocessing pipeline for MR images integrated with a baseline convolutional neural network (CNN) architecture to quantitatively assess the contribution of preprocessing to early-stage brain tumour classification. Using the Kaggle Brain MRI Dataset comprising 7023 axial T1-weighted MRI images across four diagnostic categories (glioma, meningioma, pituitary tumour, and no tumour), we implemented a multi-stage preprocessing pipeline incorporating skull stripping via morphological operations, non-local means (NLM) denoising, contrast-limited adaptive histogram equalization (CLAHE), z-score intensity normalization, and geometric augmentation (rotation ± 15°, horizontal flip, angular tilt). Both raw and preprocessed datasets were evaluated using an intentionally lightweight baseline CNN comprising two convolutional blocks (32 and 64 filters), max-pooling, dropout regularization ( p = 0.5), Adam optimizer (learning rate = 0.001), trained over 30 epochs with a stratified 70/15/15 train–validation–test split. Performance was evaluated via 5-fold cross-validation repeated across five independent random seeds (25 runs per condition), and statistical significance was assessed using McNemar’s test with 95% confidence intervals reported for all metrics. Results Preprocessing yielded statistically significant improvements across all evaluation metrics ( p < 0.001). Overall classification accuracy increased from 86.2% [95% CI 84.1–88.3] to 91.8% [95% CI 90.1–93.5] (+ 5.6% points; Wilson 95% CIs over the held-out test set). Macro-averaged precision improved from 84.1 to 90.5%, recall from 82.7 to 89.7%, and F1 score from 83.4 to 90.1%. Early-stage tumour sensitivity defined as detection of tumours occupying < 15% of brain parenchymal area improved from 73.2% [95% CI 69.8–76.6] to 85.2% [95% CI 82.4–88.0] (+ 12.0% points), rising further to 87.6% under a class-weighted loss formulation; this threshold and metric were independently validated by a panel of two radiologists on a 100-image sample (Cohen’s κ = 0.86, indicating almost-perfect inter-rater agreement), with a full per-class breakdown of qualifying images now reported (Table 4). An ablation study confirmed that the combination of skull stripping and CLAHE provided the largest marginal gains, with each preprocessing stage’s individual contribution quantified separately. Inter-class misclassification between glioma and meningioma categories decreased by 38% (raw count confusion matrices reported in full). Grad-CAM visualizations, quantified using an energy concentration ratio (ECR), confirmed that the preprocessed model produced spatially focused attention within the tumour mask (ECR = 0.84) compared to the raw model (ECR = 0.52). Multi-seed analysis confirmed that the preprocessing-driven gains were stable across random initializations. Conclusion These findings demonstrate that a well-defined preprocessing pipeline constitutes an essential and non-trivial component of deep learning-based brain tumour classification from MRI data. Given the retrospective, single-modality, JPEG-derived nature of the benchmark used, these results should be interpreted as an experimental proof-of-concept rather than evidence of clinical readiness.
Authors
- Sahil Verma (ORCID: https://orcid.org/0000-0003-3136-4029)
- Noor Zaman Jhanjhi (ORCID: https://orcid.org/0000-0001-8116-4733)
- Maninder Kaur (ORCID: https://orcid.org/0000-0003-3006-4673)
- Sayan Kumar Ray (ORCID: https://orcid.org/0000-0003-2485-437X)
- Kavita Verma (ORCID: https://orcid.org/0000-0001-5422-1659)
Institutions
- Asia University (TW)
- Rayat Bahra University (IN)
- Universidade Federal do Piauí (BR)
- Taylor's University (MY)
- Sunway University (MY)
Publication Details
- Journal
- The Egyptian Journal of Radiology and Nuclear Medicine
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1186/s43055-026-01913-x
- Primary Topic
- Brain Tumor Detection and Classification
- Type
- article
- Field-Weighted Citation Impact
- 0.00