A Novel Deep Learning Framework Utilizing Xception and CNNs for Enhancing Diagnostic Accuracy in the Precise Classification of Brain Tumors from MRI Scans

Brain tumor is regarded as one of the most aggressive diseases, affecting both children and adults. Tumors constitute 85 to 90 percent of all primary Central Nervous System (CNS) tumors. Approximately 250,000 new cases are reported annually, necessitating innovative diagnostic solutions to address the challenges associated with traditional methods. Brain tumors can lead to severe complications if not diagnosed early, making early detection crucial for effective treatment. Traditional diagnostic approaches, including manual MRI interpretation, often suffer from subjectivity and variability in expertise among radiologists, which can result in misdiagnosis and delayed treatment. This research paper presents a comprehensive novel deep-learning framework designed to classify brain tumor MRI scans into four categories: glioma, meningioma, pituitary tumor, and no tumor. Utilizing a dataset comprising 7,022 MRI images sourced from multiple repositories—including Figshare, SARTAJ, and Br35H, a novel framework is proposed by implementing a multi-task classification strategy using CNNs alongside Xception Learning to detect tumors, classify them based on malignancy and type, and identify their locations within the brain. The proposed framework achieves a training accuracy of 98.13% with a training loss of 0.1675. In comparison, validation and test accuracies were 95.11% and 93.60%, respectively with precision and recall metrics exceeding 90% for most classifications, supported by confusion matrices demonstrating strong classification capabilities. Grad-CAM techniques produce saliency maps that explain the parts affecting model predictions to understand them better. Additionally, the intended novel framework is developed via an interactive web application built via Streamlit that will allow the users to input their own MRI scans for real-time classification that will give them all the outputs, ranging from predicted tumor class up to probabilities of all the four types of tumors presented as a bar graph format along with novel saliency maps highlighting the critical areas of scans and confidence scores for each prediction. This research brings deep learning potential techniques in transforming the detection and classification of a brain tumor. It results in better achievements for the patient by obtaining prompt and precise diagnosis while overcoming the constraints of the conventional manual method.