Efficiency of Automated Brain Tumor Detection using a Deep Learning approach ResNet50 over Convolutional Neural Network models

Abstract

Brain tumors, whether non-cancerous or cancerous, can have a serious impact on brain function, potentially causing neurological problems and life-threatening complications [1]. Early and correct diagnosis of brain tumors is crucial in improving the patient prognosis and ensuring proper treatment techniques are effective. Magnetic Resonance Imaging (MRI) is among the most prevalent modalities in the diagnosis of brain tumors, considering it generates high-resolution images of soft tissues. Despite the fact that inspecting the MRI scans manually may be time-consuming, subject to human error, and heavily reliant on the level of experience of the radiologist [2].

This research studies the execution of deep learning methods that employ a highly fine-tuned ResNet50 model in the automatic brain tumor classification task. Through the utilization of transfer learning and data augmentation procedures, the model maximizes its generalization performance and ability to yield correct outputs. The model's training and validation were conducted using a set of 2,577 MRI images that were evenly divided between tumor and non-tumor cases. Model performance was improved using a variety of preprocessing and augmentation strategies. The trained model achieved a very good test accuracy of 97.35%, and high precision, recall, and F1-score, showing it is suitable for the task at hand [3] for brain tumor and non-tumor image classification.

MRI scans enable the identification of cerebral abnormalities through the presentation of both morphological and descriptive information on the size, location, and structural features of the tumor. However, traditional MRI scan analysis is typically time-consuming, human-experience-dependent, and susceptible to human error. Moreover, diagnostic readings by varying radiologists tend to be very different from each other, further raising the likelihood of misdiagnosis or late detection of a tumor. This is the justification of a necessity for computer-based diagnostic systems that would enhance the accuracy and efficiency of brain tumor detection [4], ultimately allowing for increased confidence in diagnosis and decreasing the incidence of errors. Future aims should be to promote generalizability for other MRI datasets, address real-world clinical issues, and develop hybrid AI methodologies to aid in detection capabilities [5].