ArteryPhantom-42k: A Labeled B-Mode Ultrasound Dataset for Transducer-to-Vessel Alignment Classification

ArteryPhantom-42k is a labeled B-mode ultrasound dataset for training and evaluating deep-learning models that classify ultrasound transducer-to-vessel alignment. The dataset comprises 42,266 frames extracted from 12 imaging sessions on a custom phantom: 6,612 frames labeled into three classes (center, not_center, background, balanced at 2,204 per class) plus 35,654 unlabeled frames suitable for semi-supervised and self-supervised learning. The phantom consists of seven FDM-printed TPU-90 tubes with inner diameters 7.0-10.0 mm, embedded in fresh beef shoulder tissue and imaged with a GE Logiq S7 Expert scanner and a 9L-D linear transducer on 30 January 2025. The release also includes 24 videos (12 full-FOV originals at 1474x618 plus 12 ROI crops at 958x532) and four custom Python preprocessing tools (frame extraction, ROI cropping, labeling tool, quality-control reviewer). Inter-annotator agreement on a blinded 200-frame re-label was Cohen's kappa = 0.76 (substantial agreement, 84.5% observed agreement); linearly weighted kappa_w = 0.83 (almost perfect). All 31 disagreements were single ordinal steps, concentrated at the not_center/background boundary; center vs background were never confused. The dataset is released under CC-BY-4.0; the accompanying Python tools under tools/ are released separately under the MIT License.