Deep Learning-Based Automated Detection and Classification of Brain Tumors A Case Study Implementation Using YOLOv۸ Algorithm

Background and Aims: The application of artificial intelligence (AI), particularly deep learning, in the detection and classification of brain tumors has emerged as a pivotal research domain in medical imaging. YOLOv۸, an advanced real-time object detection algorithm, provides rapid processing and high precision for identifying tumors in magnetic resonance imaging (MRI) scans. However, challenges such as variability in tumor morphology, inconsistent MRI quality, and imbalanced datasets compromise detection reliability. This study seeks to develop an automated, high-accuracy system leveraging YOLOv۸ to support medical professionals in achieving faster and more precise brain tumor diagnoses, thereby enhancing patient outcomes. Methods: A dataset comprising ۲,۱۷۵ MRI images was utilized, divided into training (۸۰%), validation (۱۵%), and testing (۵%) subsets. The YOLOv۸ single-stage architecture was implemented to simultaneously localize tumors and classify their types with minimal latency. Model training was conducted until optimal performance was achieved at iteration ۴۸, with performance assessed using accuracy and mean average precision at a ۵۰% intersection-over-union threshold (mAP۵۰). Results: The YOLOv۸-based model demonstrated exceptional detection accuracies of ۹۵% for pituitary tumors and ۹۷% for meningiomas, with the 'no tumor' class achieving ۸۷.۵% accuracy. In contrast, glioma detection exhibited the lowest accuracy at ۶۸.۲%, largely due to a high false-positive rate in this category. The model achieved an overall mAP۵۰ of ۹۴% at peak performance (iteration ۴۸), affirming its robustness in tumor localization and classification. These results underscore the need for targeted improvements in glioma detection, including larger training datasets, advanced data augmentation techniques, and hyperparameter optimization. Conclusion: The YOLOv۸-based system exhibited high accuracy in detecting pituitary tumors and meningiomas but faced challenges with gliomas, reflecting class-specific limitations. To improve performance, we recommend expanding the glioma training dataset, employing sophisticated data augmentation, fine-tuning model parameters, and reassessing classification approaches. This system holds promise as a reliable assistive tool for clinicians, offering the potential to accelerate diagnostic processes and optimize treatment planning.

کلیدواژه ها:

Brain Tumor ، Artificial Intelligence ، Deep Learning