EffUNet++: A Novel Architecture for Brain Tumor Segmentation Using FLAIR MRI Images

The advancement of emerging technologies is fundamentally reshaping the landscape of brain tumor diagnosis and treatment planning, with a notable emphasis on precise segmentation techniques to enable early intervention. Traditional manual segmentation methods encounter substantial challenges stemming from inherent noise and intensity variations in medical imaging data. Consequently, there has been a paradigm shift towards leveraging deep learning methodologies, particularly UNet++, to address these issues. The proposed methodology integrates EfficientNetB7 as the encoder component of the UNet++ architecture. This integration harnesses the power of transfer learning, specifically leveraging pre-trained weights from the AdvProp dataset, to facilitate knowledge transfer for improved performance in segmentation tasks. UNet++’s decoder module facilitates multi-scale feature fusion and enhances mask refinement through the incorporation of skip connections. The segmentation head of the architecture generates pixel-wise masks corresponding to different segmentation classes. In our experimental setup, we utilize the Kaggle lower-grade gliomas (LGG) dataset comprising 110 patient datasets for both training and evaluation purposes. The model demonstrates remarkable performance enhancements, achieving a Dice coefficient of 0.9387 and a mean Intersection over Union (IoU) of 0.9123, outperforming previous methodologies. Particularly noteworthy is the model’s ability to accurately delineate tumor boundaries and identify tumor regions, underscoring its potential for enhancing clinical decision-making and ultimately improving patient outcomes.