BMRI-NET: A Deep Stacked Ensemble Model for Multi-class Brain Tumor Classification from MRI Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Brain tumors pose a significant health risk, necessitating accurate and timely diagnosis for effective treatment and improved patient survival.
• Classifying different types of brain tumors presents a challenge due to their varied characteristics.
• While deep learning models have been applied to brain tumor classification, their accuracy requires further enhancement for clinical utility.

Purpose of the Study:

• To design and propose a novel deep stacked ensemble model, BMRI-NET, for highly accurate classification of three types of brain tumors.
• To improve the accuracy and recall of brain tumor detection from MR images compared to existing methods.

Main Methods:

• Development of BMRI-NET, a novel deep stacked ensemble model integrating three pre-trained models: DenseNe201, ResNet152V2, and InceptionResNetV2.
• Utilizing a stacking technique to combine predictions from individual models, enhancing generalization capability and overall accuracy.
• Evaluation of the model on the Figshare brain MRI dataset, comprising 3064 images of three brain tumor types.

Main Results:

• The proposed BMRI-NET model achieved an overall classification accuracy of 98.69%.
• The model demonstrated high performance with an average recall of 98.33%, F1-score of 98.40%, and MCC of 97.95%.
• Experimental results confirmed the robustness and superiority of BMRI-NET over existing brain tumor classification methods.

Conclusions:

• The BMRI-NET model offers a highly accurate and robust solution for classifying three types of brain tumors from MRI data.
• The model's superior performance indicates its potential to assist healthcare professionals in improving the accuracy and efficiency of brain tumor diagnosis.
• This advancement in deep learning for medical imaging can contribute to better patient outcomes through timely and precise treatment planning.