Deep Learning-Based MRI Brain Tumor Segmentation With EfficientNet-Enhanced UNet

Medical image segmentation plays a critical role in the field of medical image processing. Precisely delineating brain tumor areas from multimodal MRI scans is crucial for clinical diagnosis and predicting patient outcomes. However, challenges arise from similar intensity patterns, varying tumor shapes, and indistinct boundaries, which complicate brain tumor segmentation. Traditional segmentation networks like UNet face difficulties in capturing comprehensive long-range dependencies within the feature space due to the limitations of CNN receptive fields. This limitation is particularly significant in tasks requiring detailed predictions such as brain tumor segmentation. Inspired by these constraints, this study suggests incorporating EfficientNet as an encoder within UNet, with a thorough reassessment of its fundamental components: the encoder, bottleneck, and skip connections. EfficientNet replaces UNet’s encoder, initially frozen to retain learned features from pre-trained weights, adept at extracting detailed features crucial for precise segmentation like brain tumors from MRI scans. Preserving UNet’s bottleneck compresses EfficientNet’s outputs, while skip connections maintain spatial integrity during decoder upsampling. The decoder reconstructs the original image size by merging encoder-decoder features, refining boundaries with convolutional layers for accurate clinical insights. The study conducted multiclass operations on the Brain-Tumor.npy dataset from Kaggle, consisting of 3064 T1-weighted contrast-enhanced images from 233 patients with meningioma (708 slices), glioma (1426 slices), and pituitary tumor (930 slices). Experimental findings in brain tumor segmentation tasks show that the proposed model achieves performance on par with or better than recent CNN or Transformer models. Specifically, the model achieves an accuracy of 0.9925 and a loss of 0.2991 on the dataset.