Leveraging YOLOv11's Advanced Convolutional Neural Network Architecture for Improved Detection and Classification of Brain Tumors

Introduction: Recent improvements in the architectures of off-the-shelf convolutional neural networks (CNN) have enhanced their utility in medicine. You Only Look Once (YOLO) is a popular pre-trained CNN widely known for its ability to complete both object detection and classification in a single pass. This novel approach has revolutionized real-time classification, allowing predictions to be made within milliseconds on modern hardware. YOLOv11 builds upon these strong foundations by further enhancing both the speed and accuracy of classification.

Methods: 6,232 axial, sagittal, and coronal MRI cross sections were obtained from a public database and classified into four groups: no tumor, glioma, meningioma, and pituitary tumor. The largest publicly available YOLOv11 classification model (Model X) was used as a base, along with its pre-trained weights, upon which fine-tuning was performed. During the custom training process, hyperparameters such as batch size, learning rate, and optimizer settings were adjusted to optimize the model's performance. The resulting custom-trained model was then evaluated on a test set of 796 unseen cross sections (199 each) to determine its accuracy, precision, recall, and overall effectiveness.

Results: The pre-trained, highly tuned YOLOv11 model achieved an overall detection accuracy of 99%, outperforming a comparable convolutional neural network without pre-training (92% accuracy). The custom model achieved a precision and recall of 99% each for glioma detection, 98% precision and 99% recall for meningioma cases, and 100% precision and 98% recall for pituitary tumors. It perfectly classified 'no tumor' images with 100% precision and recall.

Conclusion : Exceptionally high accuracy and precision, combined with low computational requirements, have greatly expanded the potential roles of CNNs. While clinicians should always have the final say in the diagnosis of brain tumors, we propose that lightweight and accurate CNNs may provide a helpful second set of eyes, so to speak, and prevent potentially missed diagnoses due to human error.