Addressing the Impact of Resolution Scaling on YOLO Performance for Brain Tumor Detection Through Optimized Network Depth/Width Adjustments

Deep learning-based object detectors, particularly You Only Look Once (YOLO) architectures, have demonstrated strong performance in automated brain tumor detection. However, the impact of resolution scaling on tumor localization accuracy remains underexplored, especially under conditions where image resolution is reduced. This study aims to investigate how lowering the input resolution from 640 × 640 to 480 × 480 affects detection performance and whether optimized depth/width scaling and hyperparameter tuning can compensate for the expected loss of spatial detail. In this work, we propose an optimized YOLO-based framework for brain tumor detection and localization in MRI scans, building upon the method “Addressing the Impact of Resolution Scaling on YOLO Performance for Brain Tumor Detection through Optimized Network Depth/Width Adjustments.” Our model, an enhanced variant of the BGF-YOLO architecture, is specifically tailored for the challenges of medical imaging. The proposed network features both architectural and training-level optimizations. We used a publicly available dataset from Kaggle that consists of 500 training images, 201 validation images, and 100 test images. Experimental analysis demonstrates that while reducing input resolution alone degrades performance, integrating targeted modifications specifically increases network depth and width. In addition, advanced training strategies such as MixUp augmentation, dropout regularization, AdamW optimization, cosine learning rate scheduling, and finely tuned learning rate ranges lead to substantial performance gains. The optimized model achieves a precision of up to 0.858, a recall of 0.943, mAP50 of 0.946, and mAP50−95 of 0.672. These results not only outperform the reduced-resolution baseline but also approach, and in some cases surpass, the original high-resolution BGF-YOLO setup.