Performance Evaluation of U-Net Architecture for Brain Tumor Segmentation in FLAIR MRI of Lower-Grade Glioma

Abstract

Manual delineation of glioma on magnetic resonance imaging is slow, and two readers rarely produce the same contour, which weakens its value for the volumetric follow-up that lower-grade glioma management relies on. Encoder-decoder convolutional networks are the usual automatic substitute. Most of the reported improvement on this task, however, has come from larger models with pretrained backbones, and it is not obvious how much accuracy survives once the network is kept deliberately small and the training set is limited to a few thousand annotated slices. We evaluated a compact U-Net on the TCGA-LGG collection distributed through The Cancer Imaging Archive, which contains 3,939 axial FLAIR slices from 110 patients, each paired with an expert-drawn binary mask of the FLAIR abnormality. Our network uses standard 3 x 3 convolutions in a four-stage encoder and a 512-channel bridge, then switches to depthwise-separable convolutions throughout the decoder. That single substitution removes 68.1% of the decoder parameters and brings the whole model to 5.69 million weights, about one fifth the size of the original U-Net. Training used the Adam optimiser at a learning rate of 0.001 with binary cross-entropy loss, a batch size of 8 and 50 epochs, on a split of 2,836 training, 709 validation and 394 test slices. On the held-out test split the model reached an accuracy of 0.93, a sensitivity of 0.92, a specificity of 0.94, a precision of 0.88 and an F1-score of 0.90. Training and validation Dice stayed within roughly 0.02 of one another, so the model did not overfit badly on this amount of data. Tumour-free slices produced empty masks without scattered false activations. Where the model did go wrong, it went outward rather than inward: precision fell below sensitivity, and the qualitative panels put the disagreement at the infiltrative margin rather than in the lesion core, with the smallest lesion in our sample segmented least precisely. The resulting F1-score sits above the 0.84 inter-reader Dice agreement reported for this lesion type and above the 0.82 to 0.85 band reported for comparable single-pass U-Net baselines on the same collection, while staying below the 0.92 to 0.94 obtained by heavier pretrained-encoder variants. We read this as evidence that a small model recovers most of the achievable accuracy here, and that the remaining error is a boundary problem rather than a capacity problem. Because the evaluation rests on one train-validation-test partition with no cross-validation and no external test set, the results establish feasibility and not clinical readiness.