Naeem Awad Bin, Osman Onur, Alsubai Shtwai, Cevik Taner, Zaidi Abdelhamid, Rasheed Jawad
Department of Computer Science, National College of Business Administration and Economics, Multan, Pakistan.
Department of Biomedical and Neuromotor Sciences, University of Bologna, Bologna, Italy.
Front Med (Lausanne). 2025 Aug 29;12:1636059. doi: 10.3389/fmed.2025.1636059. eCollection 2025.
This study aims to develop a robust and lightweight deep learning model for early brain tumor detection using magnetic resonance imaging (MRI), particularly under constraints of limited data availability. Objective: To design a CNN-based diagnostic model that accurately classifies MRI brain scans into tumor-positive and tumor-negative categories with high clinical relevance, despite a small dataset. Methods: A five-layer CNN architecture-comprising three convolutional layers, two pooling layers, and a fully connected dense layer-was implemented using TensorFlow and TFlearn. A dataset of 189 grayscale brain MRI images was used, with balanced classes. The model was trained over 10 epochs and 202 iterations using the Adam optimizer. Evaluation metrics included accuracy, precision, recall, F1 Score, and ROC AUC.
The proposed model achieved 99% accuracy in both training and validation. Key performance metrics, including precision (98.75%), recall (99.20%), F1-score (98.87%), and ROC-AUC (0.99), affirmed the model's reliability. The loss decreased from 0.412 to near zero. A comparative analysis with a baseline TensorFlow model trained on 1,800 images showed the superior performance of the proposed model.
The results demonstrate that accurate brain tumor detection can be achieved with limited data using a carefully optimized CNN. Future work will expand datasets and integrate explainable AI for enhanced clinical integration.
本研究旨在开发一种强大且轻量级的深度学习模型，用于使用磁共振成像（MRI）进行早期脑肿瘤检测，特别是在数据可用性有限的情况下。目标：设计一种基于卷积神经网络（CNN）的诊断模型，尽管数据集较小，但能将MRI脑部扫描准确分类为肿瘤阳性和肿瘤阴性类别，且具有高度临床相关性。方法：使用TensorFlow和TFlearn实现了一个五层CNN架构，包括三个卷积层、两个池化层和一个全连接密集层。使用了一个包含189张灰度脑MRI图像的数据集，类别均衡。该模型使用Adam优化器进行了10个轮次和202次迭代的训练。评估指标包括准确率、精确率、召回率、F1分数和ROC曲线下面积（ROC AUC）。
所提出的模型在训练和验证中均达到了99%的准确率。关键性能指标，包括精确率（98.75%）、召回率（99.20%）、F1分数（98.
