tsta

فارسی

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Hybrid Active Learning–Driven Subset Dataset Selection Enables Near-Optimal Cardiac X-Ray Segmentation with Less Training Data

Authors :

Sayna Jamaati¹ Masoud Noroozi² Hossein Arabi³ Alireza Karimian⁴

1- دانشگاه صنعتی شریف 2- گروه مهندسی پزشکی، دانشکده مهندسی، دانشگاه اصفهان 3- دانشگاه ژنو 4- گروه مهندسی پزشکی، دانشکده مهندسی، دانشگاه اصفهان

Keywords :

Active Learning،Coreset Selection،Cardiac X-Ray Segmentation،Data Efficiency،Uncertainty Estimation،Deep Learning

Abstract :

Deep neural networks deliver state-of-the-art performance in medical-image segmentation but demand exhaustive pixel-level annotation, a major barrier in data-limited clinical settings. We introduce a single-shot hybrid subset-selection scheme that uses only 15% of the available training images (206 of 1,373; i.e., 85% reduction) while maintaining accuracy close to a full-data baseline. A U-Net with a ResNet-50 encoder is trained on 512×512 cardiac radiographs. To curate a compact training set, we combine (i) an uncertainty-aware active branch that converts class-probability vectors into a label-free pool-scoring uncertainty proxy via a lightweight ResNet-18 regressor, and (ii) a geometry-preserving coreset branch that applies k-center greedy search to Global-Average-Pooled (GAP) bottleneck features. The two branches are deduplicated and budgeted to 206 images (15%) before training the segmenter de novo with the same recipe as the baseline. With the fixed 15% budget, Active-only and Coreset-only training achieve Dice scores of 0.916 and 0.929, respectively; the Hybrid selection attains Dice 0.950, accuracy 0.984, sensitivity 0.950, and specificity 0.990, with no statistically significant difference from the full-data model (paired two-sided t-tests, p>0.05). The proposed mask-agnostic (at pool-scoring) and modular hybrid selector enables near-optimal segmentation at a fraction of labeling cost and is readily adaptable to other backbones and representation spaces.