Scholarlink

The brain is one of the most vital organs in the body, responsible for regulating other organ functions and enabling sound decision-making. However, various factors, including environmental influences and genetic predispositions, can trigger abnormal cell growth in the central nervous system, particularly in the brain, leading to conditions such as brain tumors. These tumors can be either benign or malignant and are further classified into specific subtypes. Traditional methods for detecting and classifying brain tumors often involve imaging techniques like MRI and CT scans, along with biopsy procedures to analyze tissue samples. While effective, these processes are time-consuming and heavily reliant on the expertise and availability of medical professionals, potentially causing delayed diagnoses and impacting patient survival due to late treatment. To address these challenges, deep learning-based solutions have been developed to provide timely and accurate diagnoses. This study focuses on fine-tuning the hybrid CNN + LSTM model proposed by Sino Cruz and Caro, initially designed for binary classification. The model was adapted and fine-tuned to perform multiclass classification of brain tumors into categories such as Meningioma, Glioma, and Pituitary tumors. The study compares the performance of the fine-tuned hybrid model using two optimizers—Adam and RMSProp—and benchmarks it against the VGG-16 model. Results indicate that the hybrid CNN + LSTM model achieved an accuracy of 83% with the Adam optimizer, outperforming its RMSProp counterpart, which achieved 78%. Interestingly, VGG-16 also achieved an accuracy of 83%, comparable to the hybrid model with Adam optimization. All models performed well in identifying Pituitary tumors and normal brain scans, but they struggled with Meningioma classification. This difficulty in recognizing Meningioma can be attributed to several factors such as the tumor’s physical property, image quality of the MRI scan, and the limited number of training set for Meningioma. Future studies could explore the potential of Transformer-based models for multiclass classification tasks. Additionally, the use of locally available hospital datasets should be considered to evaluate model performance using locally-available MRI scans. Finally, validating classification results through machine learning-based solutions using biomarkers obtained via biopsy—widely regarded as conclusive—could further enhance the reliability of these approaches.

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

Fine-Tuning Hybrid CNN + LSTM for Brain Tumor Multiclass Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Matthew John F. Sino Cruz,
  • Jaime D. L. Caro,
  • Geoffrey A. Solano

摘要

The brain is one of the most vital organs in the body, responsible for regulating other organ functions and enabling sound decision-making. However, various factors, including environmental influences and genetic predispositions, can trigger abnormal cell growth in the central nervous system, particularly in the brain, leading to conditions such as brain tumors. These tumors can be either benign or malignant and are further classified into specific subtypes. Traditional methods for detecting and classifying brain tumors often involve imaging techniques like MRI and CT scans, along with biopsy procedures to analyze tissue samples. While effective, these processes are time-consuming and heavily reliant on the expertise and availability of medical professionals, potentially causing delayed diagnoses and impacting patient survival due to late treatment. To address these challenges, deep learning-based solutions have been developed to provide timely and accurate diagnoses. This study focuses on fine-tuning the hybrid CNN + LSTM model proposed by Sino Cruz and Caro, initially designed for binary classification. The model was adapted and fine-tuned to perform multiclass classification of brain tumors into categories such as Meningioma, Glioma, and Pituitary tumors. The study compares the performance of the fine-tuned hybrid model using two optimizers—Adam and RMSProp—and benchmarks it against the VGG-16 model. Results indicate that the hybrid CNN + LSTM model achieved an accuracy of 83% with the Adam optimizer, outperforming its RMSProp counterpart, which achieved 78%. Interestingly, VGG-16 also achieved an accuracy of 83%, comparable to the hybrid model with Adam optimization. All models performed well in identifying Pituitary tumors and normal brain scans, but they struggled with Meningioma classification. This difficulty in recognizing Meningioma can be attributed to several factors such as the tumor’s physical property, image quality of the MRI scan, and the limited number of training set for Meningioma. Future studies could explore the potential of Transformer-based models for multiclass classification tasks. Additionally, the use of locally available hospital datasets should be considered to evaluate model performance using locally-available MRI scans. Finally, validating classification results through machine learning-based solutions using biomarkers obtained via biopsy—widely regarded as conclusive—could further enhance the reliability of these approaches.