Abstract Writer

Generated an IEEE-style unstructured abstract (218 words) plus a structured variant and a plain-language summary.

IEEE abstract (unstructured, 218 words)

Medical image segmentation typically requires thousands of pixel-level annotations that are expensive and slow to produce. We present LabelSparse, a semi-supervised framework that reduces annotation requirements by an order of magnitude while matching fully supervised accuracy. LabelSparse combines a consistency-regularized student-teacher network with a novel class-balanced uncertainty sampling strategy that selects the most informative unlabeled slices for pseudo-labeling. On three public benchmarks (LiTS for liver, BraTS for brain tumor, and ACDC for cardiac), LabelSparse achieves a mean Dice of 0.91 using only 5% of labels, is within 1.5 points of the fully supervised upper bound, and outperforms the strongest semi-supervised baseline by 4.2 points. Ablations show that class-balanced sampling contributes most of the gain, improving rare-class Dice by 6.8 points over random sampling. Our results demonstrate that annotation efficiency in medical imaging is limited less by model capacity than by which examples are labeled. We release code, pretrained weights, and the sampling strategy to support low-resource clinical settings.

Structured variant (for Elsevier/Springer)

Background, Methods, Results, and Conclusions, adaptable to any four-section structured format.

Plain-language summary

Training AI to find tumors in scans usually needs thousands of hand-labeled images. We built a system that needs only 5% of those labels and works almost as well, which makes high-quality medical AI possible for hospitals that cannot afford large labeling teams.

Next steps

• Run keywords-generator for the metadata that accompanies the abstract.