Hidden Flaws in Public Thyroid AI Datasets Could Inflate Diagnostic Performance

Deep learning models that appear to detect thyroid cancer with impressive accuracy may be learning far less about tumors than researchers and clinicians assume. A new audit of a widely used public thyroid imaging archive reveals contradictory labels, leaky evaluation practices, and a troubling possibility: that some of the reported diagnostic performance may come from technical artifacts in how the images were acquired rather than from any genuine sign of disease. The study, published in BMC Medical Imaging, offers a sobering case study in how easily artificial intelligence benchmarks in medical imaging can mislead.

The archive in question, released by Pang and colleagues, contains 387 thyroid ultrasound images and 1,332 fine-needle aspiration cytology blocks, and has served as a training and testing resource for numerous studies aiming to automate the distinction between benign thyroid nodules and papillary carcinoma. Behnam Kiani Kalejahi of Central Asian University in Tashkent and Mohammad Javad Rajabi of INTI International University in Malaysia set out to examine what the performance numbers reported on this archive actually demonstrate, and what they do not.

The first problem the auditors uncovered was at the level of the labels themselves. Nine ultrasound images appeared in the archive under both benign and papillary carcinoma labels, accounting for 19 files in total. In other words, the same image was simultaneously presented to learning algorithms as an example of cancer and of its absence. Such contradictory labels inject noise into training and make any reported accuracy difficult to interpret, since the ground truth against which models are judged is internally inconsistent. These conflicting cases were excluded from the audited analysis.

The second issue concerned how the data were split for evaluation. In machine learning studies, images from the same patient or the same imaging session must be kept together in either the training or the test set; otherwise the model is effectively tested on images it has already seen in near-identical form. The researchers built a frozen case-level partition, in which all images belonging to one case stayed on one side of the split, and evaluated five ImageNet-initialised convolutional network backbones, along with an ensemble averaging their probability outputs, using group-stratified five-fold cross-validation with thresholds selected on validation data and case-clustered 95 percent confidence intervals.

Under this rigorous case-level cross-validation, the cytology ensemble reached an area under the receiver operating characteristic curve, or AUROC, of 0.958, with a 95 percent confidence interval of 0.938 to 0.975, while the ultrasound ensemble achieved a more modest 0.768, with an interval of 0.716 to 0.819. But the comparison with conventional image-level splitting was revealing. When the two tested backbones were retrained using image-level folds, in which 99.8 percent of cases were contaminated across folds, the AUROC changed by only −0.002 for one backbone and +0.019 for the other. The apparent insensitivity to split design suggests that the models were not primarily memorizing individual cases, which raised a deeper question: what signal were they actually using?

To probe that question, the researchers turned to shortcut learning, the phenomenon in which neural networks exploit non-anatomical regularities, such as scanner settings, image borders, or the position of annotation overlays, instead of genuine pathological features. Using a heuristic detection procedure, they identified a sonographic sector, the fan-shaped region of the frame that contains the actual ultrasound data, and retrained all five backbones and the ensemble with that region removed. If the models were truly reading lesion morphology, performance should have collapsed. It did not. After removal of the detected sector, the ultrasound ensemble still reached an AUROC of 0.690, a paired change of −0.078 from its original score. Conversely, when only the detected sector was retained and the surround discarded, the ensemble reached 0.760, a paired change of just −0.008. Substantial discriminative signal survived in both versions of the images.

The most striking result came from a control experiment that removed the images altogether. A metadata-only model, built without any convolutional network and using only frame geometry, global intensity statistics, and border statistics, achieved an AUROC of 0.749 for ultrasound classification, with a confidence interval of 0.692 to 0.804. That figure approaches the performance of the full deep learning ensemble, which means that information about whether a nodule is benign or malignant can be substantially recovered from acquisition context alone, without ever looking at the tissue. The audit also found that the sector detection procedure fell back to a fixed central box in 24.2 percent of frames, underscoring how variable and imperfect such heuristics can be.

The researchers then tested how well their frozen ultrasound models, trained on the internal archive, transferred to external datasets they had encountered during earlier analytical development. On TN3K, assessed against its distributed benign and malignant labels, the ensemble reached an AUROC of 0.700 and over-predicted malignancy, with a calibration-in-the-large value of −0.49, indicating a systematic bias in predicted probabilities. On DDTI, the ensemble showed essentially no concordance with TI-RADS categories, scoring 0.518 with a confidence interval of 0.436 to 0.605, a value indistinguishable from chance. Transport of the internal models to new settings was therefore limited and depended heavily on which endpoint was being predicted.

The authors conclude that performance on this archive cannot be interpreted as exclusively lesion-based. Contradictory labels, partition-sensitive evaluation, and the persistence of discriminative signal in masked images and in metadata alone all point to the same conclusion: models trained on this resource may be drawing on a mixture of lesion features, acquisition artifacts, and dataset-specific regularities. Because the cytology results were strong even under case-level validation, while the ultrasound results were fragile and shortcut-sensitive, the audit also illustrates that these risks differ by modality and cannot be assumed away with a single methodological fix.

For the broader field of medical imaging AI, the study is a reminder that benchmark numbers are only as trustworthy as the data pipelines behind them. Public archives accelerate research, but they also concentrate the consequences of labeling errors and silent confounders across many downstream studies. The audit’s recommendations follow directly from its findings: reconcile labels at the case level, keep all images from a case within a single split, test whether performance survives when suspicious image regions or metadata are removed, and validate calibration as well as discrimination on external cohorts. Without such checks, a model that appears ready for the clinic may simply have learned to read the machine, not the patient.

Subject of Research: Auditing data quality and shortcut learning in deep learning models for thyroid ultrasound and cytology classification

Article Title: Case-level audit of thyroid ultrasound and cytology deep learning: label conflicts, split design, and shortcut-sensitive performance in a public archive

Article References: Case-level audit of thyroid ultrasound and cytology deep learning: label conflicts, split design, and shortcut-sensitive performance in a public archive. (n.d.). https://doi.org/10.1186/s12880-026-02944-6

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02944-6

Keywords: thyroid cancer, deep learning, ultrasound, cytology, data leakage, shortcut learning, model evaluation, reproducibility, medical imaging, papillary thyroid carcinoma, benchmark datasets, AI diagnostics