Enhancement of brain mr images and detection of brain tumors with deep learning methods

The primary aim of this study is to employ deep learning–based object detection models on magnetic resonance MR images containing brain tumors and to systematically examine the effects of different image enhancement methods on model performance. Since early and accurate diagnosis is clinically crucial, Gaussian, median, Wiener, CLAHE, and morphological operations, as well as their pairwise combinations, were applied to reduce noise, contrast irregularities, and low-quality artifacts in the images. The study was conducted using a dataset consisting of 780 MR images containing brain tumors. YOLOv8, SSD MobileNet-V2, and DETR models were tested on both raw and enhanced images. In the performance evaluation of deep learning models, mAP@0.5, mAP@[0.5:0.95], precision, and recall metrics were employed, whereas in the image enhancement stage, MSE, PSNR, and SSIM metrics were utilized. In particular, Gaussian filtering significantly reduced MSE values and improved PSNR and SSIM metrics, leading to a clear enhancement in image quality. The results indicate that image enhancement provided a notably positive contribution, especially for the SSD model, while its effects on YOLOv8 and DETR models were more limited or varied depending on the method applied. The SSD model achieved the highest performance gain, with a 12% increase in confidence scores and approximately a 9.8% improvement in mAP after enhancement. The Gaussian–Closing combination yielded the best results for the DETR model, whereas CLAHE–Dilation led to performance degradation in all three deep learning models. Overall, combining blurring and morphological operations resulted in an approximate 10% increase in confidence scores for SSD and produced more limited but measurable performance variations for YOLOv8 and DETR depending on the applied method. These findings highlight that image enhancement strategies should be evaluated in conjunction with the architectural characteristics of the deep learning models used.