Paper Reviews

Resampling Uncertainty Estimation asks a narrower, more auditable question than most uncertainty methods: not how confident is the model in general, but how much would this specific prediction have changed if the model had been fit on slightly different training data.

MONET learns a dense, concept-scoring image-text space from 105,550 dermatology images paired with literature descriptions, then uses that space for data auditing, model auditing, and interpretation, including recovering a spurious redness-malignancy correlation that flipped sign between two hospitals.

Progressive stress testing with five perturbation types at six severity levels, applied to chest X-ray and skin-lesion classifiers, separates iid test accuracy from robustness and finds that pretraining choice, not just architecture, shapes how a model degrades.

A skin-lesion classifier at dermatologist-level AUC gave false positive or negative predictions for up to 22% of lesions under a simple image rotation, a direct test of whether an acceptable benchmark score means the model is ready for a clinic.

TCAV assumes a concept occupies a linear direction in activation space. Concept Gradients drops that assumption and, in a medical case study, tracks concept importance scores against mortality-risk descriptions already published in the clinical literature.

A viewpoint proposing a structured audit process for deployed medical AI, treating shortcut learning, poor generalizability, and weak explainability as the three named failure modes an auditor should anticipate rather than discover by accident.

Generative counterfactuals read by two blinded dermatologists across five academic and consumer skin-lesion classifiers, the closest published template I have found for the audit I want to run on ultrasound models.

A framework built around the fact that a single-grade disagreement on the KL scale can redirect a patient from conservative therapy to a surgical pathway, explainability motivated directly by clinical stakes, not as a generic add-on.

A foundation model trained on 0.87 million image-report pairs, built from the ground up around clinical-concept embeddings rather than bolting interpretability onto a black box afterward, auditability as a design constraint, not an add-on.

A multitask model that predicts BI-RADS descriptors alongside malignancy, one of the earlier attempts at the same explainable-by-BI-RADS idea that later concept-bottleneck papers pushed further.

A method for detecting shortcut reliance by looking directly at correlations in a network's weight space, rather than at its activations or explanations, a different vantage point on the same underlying question.

The ICDR scale: a 31-person, 16-country consensus process behind the DR severity labels nearly every retinal AI paper trains against, built via a modified Delphi method rather than a single objective measurement.

A tone-mapping method that expands ultrasound's usable dynamic range before a model, or a clinician, ever sees the image, quantified with real, measurable image-quality gains.

A counterfactual-generation approach to quantifying shortcut reliance in disease classifiers, in the same spirit as RoentMod but framed as a general evaluation method rather than a modality-specific tool.

A McIntosh-lab paper on fixing shortcuts at the data level (laterality markers, scanner noise) without the mitigation itself quietly introducing a new confounder, which the authors show is a real risk of naive data-level fixes.

Rather than predicting a single aggregate KL grade, this model grades each individual OARSI-atlas feature (osteophytes, joint space narrowing, and more) separately, matching the atlas's own granularity.

The Rashomon-set argument: for a given dataset, there are often many equally accurate models, some far simpler and more interpretable than others, so 'we needed the complex model for accuracy' is a claim that should be checked, not assumed.

A study showing that models can achieve high internal accuracy by exploiting signals tied to how and where data were collected, and lose that performance when transferred elsewhere.

CDEP, penalizing a model's contextual-decomposition-based explanation directly during training, so an explanation isn't just a diagnostic afterward but an actionable lever to fix the model.

An early, template-based method for finding microaneurysms, the first and smallest lesions of diabetic retinopathy, using wavelet-domain matching rather than a learned classifier.

A framework connecting interpretability directly to correction: find the spurious behavior with concept-level explanation methods, then fix both the model and the data it came from.

A CAD system targeted specifically at the ambiguous BI-RADS category 3, 'probably benign' masses, testing whether quantitative features can safely reclassify malignant cases that radiologists had grouped as low-risk.

A pre-deep-learning digital measurement system for knee-OA radiographic features, a reminder that 'objective, automated quantification' of joint space and osteophytes predates neural networks by well over a decade.

A frequency-domain attention module built to recover the high-frequency texture and boundary information that Vision Transformer segmentation models tend to lose, directly relevant to any medical target defined by fine texture or a sharp boundary.

A causal framing of the shortcut-learning problem in disease diagnosis, trying to explicitly align a model's learned representation with the causal structure of disease, not just penalize known spurious features after the fact.

A method for locating which layer of a network a shortcut's features actually manifest in, using Prediction Depth and KL divergence, moving from 'the model has a shortcut' to 'here is where in the network it lives.'

A statistics paper on adjusting rank correlation for confounding covariates, read because so many AI-versus-classical-index comparisons in medical imaging report raw Spearman correlation without asking whether a shared confound is inflating it.

Radiology-specific saliency evaluation: sanity checks and human-alignment tests applied directly to chest radiograph localization tasks.

The clinical practice guideline that defines how diabetic retinopathy is actually meant to be screened, staged, and managed, the standard any AI screening tool is ultimately deployed to support, not replace.

Joint grading of two correlated diabetic-eye diseases via cross-disease attention, a structural bet that shared retinal evidence should inform both diagnoses at once, rather than training two separate classifiers that never talk to each other.

A pre-deep-learning approach to breast-ultrasound CAD: hand-engineered wavelet channel-energy features meant to capture how infiltrative a lesion's margin looks, a useful reminder of what feature engineering used to make explicit.

A sobering stress test of the exact interpretability toolkit I lean on elsewhere in this collection (feature attribution, concept activation, and training-point ranking) against spurious signals the practitioner doesn't already know to look for.

Two models with identical training accuracy, identical architecture, and different random seeds can behave completely differently under distribution shift, a Google-scale audit of just how common this is, including in a dermatology model.

A clinical asthma-and-pneumonia case study opens this paper: a model that learned asthma predicts lower readmission risk, backwards from reality, because of how the training data was collected, and a method to penalize a model for explaining itself that way.

Modern networks are more accurate and less calibrated than their predecessors, and the fix (temperature scaling) is almost embarrassingly simple.

An extension of TCAV to continuous, graded concepts: used to show that nuclei texture, not just its presence, is a directional driver of tumor-grade predictions in lymph-node histopathology.

A head-to-head comparison of two classical ultrasound-based liver-fat indices in living liver-donor candidates, against biopsy, the pair of legible baselines that any liver-steatosis deep model is implicitly competing with.

A biopsy-proven, multi-scanner breast ultrasound dataset with standardized cross-validation partitions, the kind of dataset-hygiene paper that makes fair benchmark comparisons possible in the first place.

The saliency method almost every medical-imaging paper cites, and a good place to be precise about what its heatmap actually certifies.

A standards-body review on turning ultrasound attenuation and backscatter into a quantitative, reproducible liver-fat biomarker, the standardization work that any AI liver-steatosis model eventually has to sit on top of.

The internal grading manual behind the EyePACS diabetic-retinopathy dataset, the lesion-by-lesion rulebook that every EyePACS-trained DR model's labels ultimately trace back to.

Chest X-ray classifiers assigned 'no finding' disproportionately to underserved patients, the error mode that silently denies care.

A counterfactual image-editing model that inserts or removes a chest X-ray finding without retraining the classifier under test: used to catch models exploiting institutional, demographic, and device markers as shortcuts.

The paper that named and formalized the concept bottleneck: asking, using the paper's own example, whether a model would still predict severe arthritis if it didn't think there was a bone spur.

ResU-Net segmentation feeding 64-point multi-JSW measurements into XGBoost, outperforming the conventional single minimum-JSW pipeline for predicting knee-OA severity and progression, with the segmentation itself validated against a radiologist.

An empirical look at how well individual BI-RADS sonographic descriptors actually predict malignancy, the kind of per-feature grounding that a BI-RADS-based concept bottleneck for breast ultrasound is implicitly relying on.

Within a labeled class there are clinically distinct subsets, and a model can fail on the dangerous ones while the headline number stays high.

A transfer-learning pipeline for grading liver steatosis from B-mode ultrasound, benchmarked directly against the hepatorenal index and GLCM texture features, with biopsy as the reference standard.

During a pandemic, dozens of high-accuracy COVID detectors were published. This paper showed many were reading laterality markers, patient positioning, and dataset provenance.

A concept bottleneck that predicts BI-RADS features before the final cancer classification, and shows that letting a radiologist correct the intermediate concepts actually improves the final diagnosis.

A short clinical primer on the KL grading scale, the single most common label target for knee-OA deep learning papers, and a reminder of how much subjectivity that five-point scale is quietly absorbing.

A method (ShorT) for directly testing whether a clinical model is using shortcut correlations, with the uncomfortable finding that shortcuts are not always the reason a model is unfair.

CNNs detected the treating institution from chest radiographs and used it as a proxy for disease prevalence, an early, concrete demonstration of confounding by site.

A hand-crafted, semi-quantitative ultrasound score (US-FLI) built from four visual features, the kind of clinically legible scoring system that a learned model in the same space should be able to match or explain, not just outperform.

TCAV asks whether a model's prediction is sensitive to a human-named concept, such as 'striped' or 'malignant,' rather than to a single pixel, the tool that later work on shortcut detection in medical imaging builds directly on.

The OARSI atlas that turns joint-space narrowing, osteophytes, and sclerosis into graded reference images, the ground truth that any automated knee-OA grader is ultimately trying to reproduce.

The paper that gave the field a shared vocabulary for models that solve the benchmark without solving the task.