Scholarlink

A brain tumor is a growth of abnormal cells in the brain. As mentioned by the American Association of Neurological Surgeons (AANS), there are around 150 types of brain tumors. Medical experts can provide the proper diagnosis and treatment for patients if they have accurate information about the type and grade of the tumor. This study proposes a simple and efficient method based on deep learning and a self-supervised learning approach for brain tumor classification on a magnetic resonance imaging (MRI) dataset from the Kaggle repository. Firstly, the brain tumor dataset is processed to remove small regions of noise, crop the black background, and then resize the MRI images to the common size of 224 × 224. Secondly, the pre-trained ResNet50 model is trained with the preprocessed MRI dataset to classify three classes of brain tumors (i.e., glioma, meningioma, and pituitary) and one class of healthy (i.e., notumor). Finally, a small ResNet architecture-based network (namely ResNet_s model) from scratch and the pre-trained ResNet18 are taught by the ResNet50 model using self-supervision. The ResNet_s and ResNet18 model are almost 22x (1,083,620 parameters) and 2x (11,178,564) smaller than ResNet50 (23,516,228 parameters). The AdamW optimizer with an initial learning rate of 1e−3 and the same data augmentation techniques are applied for both transfer learning and self-supervised learning. After fine-tuning the brain tumor dataset, the ResNet50 model obtained F1 score of 0.994 on the testing set. Meanwhile, based on the self-supervised training process with a teacher of the trained ResNet50 and without any labels, the ResNet_s and ResNet18 models achieved F1 scores of 0.954 and 0.989 on the testing set, respectively. The performance of the ResNet_s and ResNet18 model shows the potential of the proposed self-supervised learning technique when it can build small (student) networks to obtain knowledge of teacher simply. In addition, this technique is different from transfer learning in that the student model can learn with unlabeled data. Besides, the small models have the ability to support the different low-power devices.

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

Brain Tumor Classification Using Deep Learning and Self-supervised Learning Approach

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Thi-Thu-Hien Pham,
  • Thanh-Hai Le

摘要

A brain tumor is a growth of abnormal cells in the brain. As mentioned by the American Association of Neurological Surgeons (AANS), there are around 150 types of brain tumors. Medical experts can provide the proper diagnosis and treatment for patients if they have accurate information about the type and grade of the tumor. This study proposes a simple and efficient method based on deep learning and a self-supervised learning approach for brain tumor classification on a magnetic resonance imaging (MRI) dataset from the Kaggle repository. Firstly, the brain tumor dataset is processed to remove small regions of noise, crop the black background, and then resize the MRI images to the common size of 224 × 224. Secondly, the pre-trained ResNet50 model is trained with the preprocessed MRI dataset to classify three classes of brain tumors (i.e., glioma, meningioma, and pituitary) and one class of healthy (i.e., notumor). Finally, a small ResNet architecture-based network (namely ResNet_s model) from scratch and the pre-trained ResNet18 are taught by the ResNet50 model using self-supervision. The ResNet_s and ResNet18 model are almost 22x (1,083,620 parameters) and 2x (11,178,564) smaller than ResNet50 (23,516,228 parameters). The AdamW optimizer with an initial learning rate of 1e−3 and the same data augmentation techniques are applied for both transfer learning and self-supervised learning. After fine-tuning the brain tumor dataset, the ResNet50 model obtained F1 score of 0.994 on the testing set. Meanwhile, based on the self-supervised training process with a teacher of the trained ResNet50 and without any labels, the ResNet_s and ResNet18 models achieved F1 scores of 0.954 and 0.989 on the testing set, respectively. The performance of the ResNet_s and ResNet18 model shows the potential of the proposed self-supervised learning technique when it can build small (student) networks to obtain knowledge of teacher simply. In addition, this technique is different from transfer learning in that the student model can learn with unlabeled data. Besides, the small models have the ability to support the different low-power devices.