Native image geometry and cross-partition patient overlap in a checksum-identified public liver ultrasound release: an audit of SMC-LUD

Abstract Background Patient-level separation and control of acquisition-related confounding are important for interpreting ultrasound artificial-intelligence benchmarks. Image-level partitioning can expose models to patient-specific information, while class-dependent image geometry can encode labels without clinical content. We performed a checksum-identified release audit of the public Samsung Medical Center Liver Ultrasound Dataset (SMC-LUD), focusing on release characteristics relevant to hepatocellular carcinoma (HCC) versus hemangioma classification. Methods We reconciled the descriptor, repository metadata, downloadable folders, and an audit-derived patient manifest for Figshare article 31112716, file 61666681 (MD5 03b28af94e37c4ae9b27024654aaf91e). We audited counts, cross-partition patient representation, hashes, native dimensions, and file format. A non-pixel geometry rule classified square versus non-square images, and frozen ImageNet embeddings retrieved the nearest training frames. Exploratory ResNet-18 analyses used released and patient-exclusive partitions; an additional post hoc sensitivity analysis converted decoded frames to three-channel grayscale on a common 224 × 224 grid. Confidence intervals used 2,000 patient-level bootstrap samples. Results The descriptor reports 5,385 images from 1,021 patients; the archive contained 5,385 readable, hash-unique images under 1,013 patient directories. All 2,716 HCC images were square, whereas none of 2,669 hemangioma images were square; native geometry classified diagnosis with 100% sensitivity and specificity. In the all-image release, 643/1,013 patients (63.5%) were represented in multiple partitions and 344/360 test patients (95.6%) were also represented in development; the latter proportion was 162/194 (83.5%) for clean images. Frozen embeddings retrieved the same patient first for 44.8% of eligible test frames (95% CI 37.7–52.0%; random expectation 0.36%). Patient-exclusive ensemble AUROC was 0.966 (95% CI 0.909–1.000) at baseline and 1.000 (95% CI 1.000–1.000) after forced-square grayscale standardization; paired difference 0.034 (95% CI 0.000–0.090). Conclusions For the checksum-identified SMC-LUD release, observed patient representation across partitions and archive-level patient counts were not fully aligned with the published description, and native geometry perfectly separated the diagnostic classes. High discrimination after patient-exclusive splitting and simple input standardization did not resolve this non-identifiable technical confounding. Public medical-imaging datasets can strengthen interpretability by providing immutable patient-exclusive manifests, harmonizing technical properties across classes, and applying automated release checks before use as benchmarks.