YOLO-LS: a novel deep learning framework for brain tumor segmentation in Magnetic Resonance Imaging

Brain tumors exhibit high heterogeneity in morphology, texture, and location, making accurate recognition and segmentation critical for clinical diagnosis, surgical planning, and prognosis evaluation. However, manual annotation of MRI scans is hindered by subjective bias and inefficiency. Furthermore, existing automated frameworks often face a trade-off between segmentation precision-particularly at infiltrative boundaries-and the computational efficiency required for deployment in resource-constrained environments. To address these challenges, this study proposes YOLO-LS (Lightweight Segmentation), an enhanced framework based on the YOLO11n-seg architecture designed for efficient detection and high-precision segmentation of brain tumors. The methodology introduces three key innovations: (1) integrating ShuffleNet V1 as a lightweight backbone to significantly reduce parameter count and computational complexity via pointwise grouped convolutions; (2) incorporating the DySample dynamic upsampling mechanism to mitigate the loss of fine-grained semantic details inherent in traditional interpolation, thereby improving the recovery of tumor boundaries; and (3) optimizing the neck network with a C3k2-PoolingFormer module to facilitate efficient cross-scale feature fusion and global context capture. The model was trained and tested on the Figshare dataset (3,064 images) using five-fold cross-validation and externally validated on an independent Kaggle dataset (300 images). Results demonstrate that YOLO-LS achieved a bounding box mAP50 of 0.953 ± 0.011, a Dice coefficient of 0.91 ± 0.01, and a 95% Hausdorff Distance (HD₉₅) of 4.35 ± 0.34 mm on the internal test set, the latter indicating superior boundary adherence compared to baseline models (paired t-test, p < 0.05 for all key metrics). Notably, the model reduced computational cost to 8.1 GFLOPs (a 15.6% reduction) while achieving an increase of 2.9% points in mAP50 compared to the baseline YOLO11n-seg. Comparative analyses with state-of-the-art architectures, including U-Net, SegNet, Swin-UNet, and VM-UNet, confirmed the efficacy of these improvements. Furthermore, Grad-CAM heatmaps validated the model's precise focus on tumor core and edge regions. On the external test set, the model exhibited strong generalization with a Dice coefficient of 0.895. These findings indicate that YOLO-LS achieves an effective balance between precision, efficiency, and interpretability, demonstrating significant potential for assisting diagnostic workflows in resource-constrained clinical environments.