An Anatomy-Aware EfficientNetB0 and CBAM Framework for Suppressing Hard Boundary Bias in MRI Brain Tumor Classification

Authors

• Ali Abdulzahraa Jalil, University of AlKafeel, Najaf, Iraq
• Ahmad Fatlawi, University of AlKafeel, Najaf, Iraq
• Hussein Ali Hussein Al Naffakh, Department of Medical Laboratory Techniques, University of AlKafeel, Najaf, IraqFollow
• Umar Farooq Khattak, UNITAR International University, Kuala Lumpur, Malaysia

How to Cite This Article

Jalil, Ali Abdulzahraa; Fatlawi, Ahmad; Al Naffakh, Hussein Ali Hussein; and Khattak, Umar Farooq (2026) "An Anatomy-Aware EfficientNetB0 and CBAM Framework for Suppressing Hard Boundary Bias in MRI Brain Tumor Classification," Polytechnic Journal: Vol. 16: Iss. 2, Article 9.
DOI: https://doi.org/10.59341/2707-7799.1883

Document Type

Original Article

Abstract

The accurate classification of brain tumors from Magnetic Resonance Imaging (MRI) is important for prognosis and treatment planning in neuro-oncology. Although Convolutional Neural Networks (CNNs) dominate medical image analysis and achieve high scores, many rely on a clinically dangerous shortcut that we call Hard Boundary Bias (HBB). When an MRI slice is saved with zero padding, the boundary between the bright skull and the dark background forms the strongest gradient in the image, and the network fixes on this artificial edge instead of the tumor. To address this, we propose a lightweight framework that combines an EfficientNetB0 backbone with a Convolutional Block Attention Module (CBAM) to build an anatomy-aware classifier. A deterministic preprocessing stage identifies the cranial region using Otsu thresholding and crops a tight bounding box around it, removing the padded background before any feature is learned and treating the bias at its source. Using 5-fold cross-validation, the model reached 97.26% ± 0.34% on the Sartaj dataset at the slice level, since no patient identifiers are distributed with this benchmark. On the Figshare dataset it reached 80.12% ± 1.38% under a strict patient-level GroupKFold protocol that prevents data leakage, with mean AUCs of 0.9967 and 0.9527. A naive slice-level split of Figshare raises the accuracy to 92.66%, and this 12.5-point gap quantifies the optimism introduced by patient-level leakage. Grad-CAM confirms that the model focuses on pathological tissue rather than background cues. Because the framework prioritizes leakage-free evaluation and interpretability over headline accuracy, we present it as a promising research prototype whose clinical reliability would require external and prospective validation.

Receive Date

18 July 2026

Revise Date

22 Aug 2026

Accept Date

22 Aug 2026

Publication Date

2026