Comparative evaluation of convolutional and transformer-based object detection models for brain tumor localization in MRI

There is a need to detect and precisely localize brain tumors based on medical imaging to ensure such tumors are diagnosed early and treated with the aim of a successful treatment. More recent developments in the field of deep learning have presented convolution-based and transformer-based object detection models that can be used to achieve high detection rates. We carry out an analogous comparison of YOLOv11s and RF-DETR(small) in brain tumor detection on annotated Magnetic Resonance Imaging (MRI) data on Roboflow in this study. Precision, Recall and mean Average Precision (mAP) along with localization loss measures were used to train and test both models. Findings indicate that YOLOv11s performed better in terms of mAP@50 of 95.2 percent, precision of 93.6 percent, and recall of 91 percent compared to RF-DETR that reported mAP@50 of 94.3 percent, precision of 90.6 percent, and recall of 85 percent. The findings point to the improved performance of YOLOv11s to localize and classify tumor locations in varying MRI slices. YOLOv11s with less convergent dynamic training behavior and RF-DETR(small) with better localization based on more challenging Intersection over Union (IoU) thresholds. The primary challenge was ensuring a fair comparison between two fundamentally different detection paradigms that employ distinct feature-learning mechanisms, pretraining schemes, optimization settings, and computational characteristics. Despite these, our findings highlight the feasibility of integrating advanced object detection models into computer-aided diagnostic tools for robust brain tumor detection in complex MRI data.