Brain tumor classification from contrast-enhanced T1 MRI, a patient-wise deep learning study
Year: 2026
Type: article
Abstract: Accurate classification of intracranial tumors on contrast-enhanced T1-weighted MRI (CE-T1) is challenged by heterogeneous enhancement patterns and strong correlations among multiple slices from the same patient. We propose a compact CE-T1 MRI analysis framework that standardizes image intensities, enhances local CLAHE (contrast using contrast-limited adaptive histogram equalization), and applies a lightweight convolutional classifier whose slice-level probabilities are aggregated to obtain a single patient-level prediction. An independent optional U-Net-based pathway provides binary tumor segmentation when spatial delineation is required. Evaluation was performed using a strictly patient-wise, leakage-controlled protocol with stratified five-fold cross-validation on 233 subjects (3,064 slices) comprising glioma, meningioma, and pituitary adenoma. The classifier achieved 99.4% patient-level accuracy, with macro-averaged sensitivity and specificity of 99.3% and 99.0%, respectively. The optional segmentation pathway achieved a mean Dice similarity coefficient of 97.3%. Post-hoc temperature scaling reduced the patient-level expected calibration error from 3.0% to approximately 1.2%, although larger independent cohorts are required to determine the stability of this estimate. The classifier contains 7.8 million parameters and requires approximately 15.9 GFLOPs per $$256\times 256$$ slice, with an end-to-end inference latency of approximately 0.41 s per slice under the evaluated hardware configuration. These findings demonstrate strong patient-level performance within the evaluated retrospective cohort. The framework is intended as a research-stage approach for reproducible evaluation rather than a clinically validated diagnostic system for routine radiology practice. However, no independent quantitative external validation was performed; therefore, multi-centre evaluation across different scanners, acquisition protocols, institutions, and patient populations is required before generalizability or clinical applicability can be established.
Source: Scientific Reports
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
Funder: Najran University
Award: G9001389880
OpenAlex ID: W7233789627
