ScintiGround-38K: A Grounded Visual Question Answering Benchmark for Paired Anterior–Posterior Bone Scintigraphy

<title>Abstract</title> <p>Whole-body bone scintigraphy (WBBS) acquires paired anterior and posterior views of skeletal tracer uptake, yet few question-answering benchmarks retain the labels and bounding boxes of their upstream data rather than reducing them to free text that cannot be checked against the image. We present ScintiGround-38K, a frozen, provenance-preserving resource for visual question answering (VQA), image localization, and grounded VQA. It contains 38,469 instances, comprising 13,371 VQA, 12,549 image-localization, and 12,549 grounded instances derived from 54,768 source records and 23,967 anterior–posterior image pairs from 2,815 patients. All bounding boxes and answer labels originate in the BS-80K source dataset, and question wording is drawn from fixed rule-based templates. Template matching recovers bone-region boxes in native image coordinates, source hotspot boxes provide lesion locations, and metadata-only rows are excluded from training and evaluation. Every instance retains its source identifier, view, native-image box, and patient-level split assignment, and raw upstream images are not redistributed. Two vision-language models, Qwen3-VL-2B-Instruct and PaliGemma-3B-pt-224, were adapted with low-rank adaptation (LoRA) on an identical multitask training set and evaluated on the held-out test set after model selection on validation data. Raw answer accuracy exceeded 0.90, yet balanced accuracy reached only 0.528 and 0.535 for visual question answering, abnormal-class F1 remained below 0.15, and strict grounded accuracy reached 0.765 and 0.701. High raw accuracy therefore conceals weak minority-class and evidence-level behaviour. ScintiGround-38K is intended for methodological research and has not undergone clinical validation.</p>