Scholarlink

<p>Brain tumors are serious and abnormal growth within the brain by posing significant challenges for medical diagnosis and treatment. The varied and complex nature of brain tumors complicates their detection and accurate classification. The need for precise classification of brain tumors from medical imaging is crucial for effective medical intervention. However, traditional methods are reliant on manual diagnosis, which are prone to human errors and inconsistencies. Previous research has struggled to enhance diagnostic precision by combining predictions from multiple models but often result in discrepancies and inconsistent outcomes by leading to complicated diagnostic processes. To address these challenges, researchers have increasingly turned to advanced computational methods to improve diagnostic accuracy. In this context, we propose a robust ensemble approach optimized using a Differential Evolution (DE)-based algorithm. Our method combines three high-performing pre-trained CNN models like MobileNetV1, MobileNetV2, ResNet50V2 and optimizes their contributions by assigning optimal weights through DE. This technique intelligently adjusts the weights allocation of models to maximize ensemble performance. During optimization, the probabilities from each model are extracted and integrated using a weighted average aggregation scheme, enhancing the diagnostic precision and overall predictive accuracy. To validate the effectiveness of our approach, we applied it to two publicly available datasets: a binary classification dataset (BR35H) and a multi-class (4-class) dataset. Through rigorous evaluations, our optimized ensemble approach demonstrated superior accuracy performance of 98% and 97.03%, respectively. We used several performance evaluation metrices and visualization techniques like Grad-CAM to highlight critical areas within the images. Additionally, statistical validation is conducted using the Friedman test followed by the Conover post hoc analysis to rigorously assess and compare the performance differences across models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Differential evolution-driven optimized ensemble network for brain tumor detection

  • 期刊论文

  • 发表日期：

    2025年4月16日

查看全文
  • Arash Hekmat,
  • Zhang Zuping,
  • Omair Bilal,
  • Saif Ur Rehman Khan

摘要

Brain tumors are serious and abnormal growth within the brain by posing significant challenges for medical diagnosis and treatment. The varied and complex nature of brain tumors complicates their detection and accurate classification. The need for precise classification of brain tumors from medical imaging is crucial for effective medical intervention. However, traditional methods are reliant on manual diagnosis, which are prone to human errors and inconsistencies. Previous research has struggled to enhance diagnostic precision by combining predictions from multiple models but often result in discrepancies and inconsistent outcomes by leading to complicated diagnostic processes. To address these challenges, researchers have increasingly turned to advanced computational methods to improve diagnostic accuracy. In this context, we propose a robust ensemble approach optimized using a Differential Evolution (DE)-based algorithm. Our method combines three high-performing pre-trained CNN models like MobileNetV1, MobileNetV2, ResNet50V2 and optimizes their contributions by assigning optimal weights through DE. This technique intelligently adjusts the weights allocation of models to maximize ensemble performance. During optimization, the probabilities from each model are extracted and integrated using a weighted average aggregation scheme, enhancing the diagnostic precision and overall predictive accuracy. To validate the effectiveness of our approach, we applied it to two publicly available datasets: a binary classification dataset (BR35H) and a multi-class (4-class) dataset. Through rigorous evaluations, our optimized ensemble approach demonstrated superior accuracy performance of 98% and 97.03%, respectively. We used several performance evaluation metrices and visualization techniques like Grad-CAM to highlight critical areas within the images. Additionally, statistical validation is conducted using the Friedman test followed by the Conover post hoc analysis to rigorously assess and compare the performance differences across models.