The clinical motivation is familiar to breast specialists: when imaging findings are ambiguous, such as in lesions classified as BI-RADS 4, or when radiologic and pathologic results are discordant, deciding whether a lesion is benign or malignant can be difficult. The team, comprising Xiaoyan Zhao, Xingxing Wu and Dan Zhao of Shanxi Provincial Children’s Hospital together with Hui Feng of Taiyuan Maternal and Child Health Hospital, built a tissue-based molecular tool intended to complement, not replace, standard histopathological diagnosis in these challenging cases.

The researchers first mined public microRNA and messenger RNA datasets to discover and prioritize candidate inverse-expression pairs, using those datasets primarily for tumor-versus-normal discovery, while reserving the critical benign-versus-malignant question for their own independently collected tissue cohort. That cohort comprised 150 breast-lesion samples, 68 benign and 82 malignant, in which the team measured ten candidate axes using reverse-transcription quantitative polymerase chain reaction, a laboratory technique that quantifies specific RNA molecules.

Two candidate pairs emerged with the strongest evidence of inverse expression: miR-205-5p paired with HMGB3, and miR-96-5p paired with FOXO1. The study’s contribution lies in these measured expression relationships within a diagnostic context rather than in new functional claims about the genes themselves.

To convert the molecular measurements into a diagnostic tool, the researchers developed a least absolute shrinkage and selection operator, or LASSO, logistic-regression model in a training cohort. The final model retained five markers: miR-205-5p, HMGB3, miR-96-5p, FOXO1 and miR-21-5p.

The performance figures come from an internal-validation design with a notable safeguard. The model was tested in a separate internal-validation cohort without any marker reselection, coefficient refitting or threshold re-optimization, meaning it faced the validation data exactly as it left the training phase. In that validation set, the area under the receiver operating characteristic curve, a standard summary of diagnostic discrimination, was 0.894, with a 95 percent confidence interval of 0.788 to 0.975. Accuracy reached 0.844, sensitivity 0.840 and specificity 0.850.

These numbers indicate that the model correctly identified roughly 84 percent of malignant lesions and 85 percent of benign ones in the held-out samples. In a diagnostically uncertain setting, a molecular adjunct with this level of discrimination could add confidence to clinical decisions. The authors frame the model as adjunctive throughout: it is meant to sit alongside imaging and pathology, not to replace them.

The study’s limitations are stated plainly. The validation is internal, and external validation in independent cohorts, ideally from multiple centers, as well as functional confirmation of the candidate miRNA-mRNA axes, are required before any clinical use can be considered. For now, the finding stands as a promising early step, and whether the miR-205-5p/HMGB3 and miR-96-5p/FOXO1 axes prove useful in the broader population of ambiguous breast lesions will depend on the external studies the authors themselves call for.

Subject of Research: A tissue-based inverse miRNA–mRNA diagnostic model for distinguishing benign from malignant breast lesions

Article Title: Identification of candidate inverse miRNA–mRNA pairs and an internally validated tissue-based diagnostic model for benign and malignant breast lesions

Article References: Zhao, X., Feng, H., Wu, X., & Zhao, D. (2026). Identification of candidate inverse miRNA–mRNA pairs and an internally validated tissue-based diagnostic model for benign and malignant breast lesions. Frontiers in Molecular Biosciences, 13, Article 1952773. https://doi.org/10.3389/fmolb.2026.1952773

Image Credits: AI Generated

DOI: 10.3389/fmolb.2026.1952773

Keywords: breast cancer, benign breast lesions, miRNA, mRNA, RT-qPCR, LASSO logistic regression, diagnostic model, miR-205-5p, HMGB3, molecular diagnostics, BI-RADS 4, Identification

The technique at the heart of the study is called radiomics, a branch of medical imaging analysis that converts ordinary pictures into vast troves of quantitative data. Where a radiologist sees a tumor’s size, shape, and density, a radiomics pipeline extracts hundreds of mathematical descriptors: measures of texture homogeneity, gray-level co-occurrence patterns, edge sharpness, and spatial complexity. The premise is seductive. Tumors that are biologically primed to die under radiation may carry telltale signatures in their pixel-level architecture, signatures that could be captured before a single treatment session begins. If such signatures could be reliably decoded, oncologists could escalate therapy for predicted non-responders, spare frail elderly patients the burden of ineffective treatment, and personalize care in ways currently impossible.

The research team retrospectively reviewed 189 patients with esophageal cancer who had undergone definitive radiotherapy at their single center. All patients had baseline contrast-enhanced CT scans from which radiomic features were extracted. The cohort divided into 126 responders and 63 non-responders based on early radiological assessment of treatment effect. Following international reporting standards known as TRIPOD, the investigators pre-specified their primary analysis: a LASSO-penalized logistic regression model, a statistical technique that shrinks the influence of weak predictors toward zero and selects only the most informative features. Crucially, they capped the number of predictors at five, a constraint dictated by the number of non-responders in the cohort, since reliable model fitting generally demands at least ten outcome events per predictor variable.

What followed is a masterclass in methodological honesty, and it is here that the study earns its place in the ongoing debate about artificial intelligence in medicine. When the team evaluated their five-feature radiomic model using fully nested cross-validation, a procedure in which every step of the analysis, from initial filtering through feature selection to final model fitting, is re-executed from scratch within each testing fold, the model’s discrimination collapsed to an area under the curve of just 0.554, with a standard deviation of 0.108. An AUC of 0.5 represents performance no better than a coin flip. In other words, when the researchers refused to let any information from the test data leak into the training process, the radiomic model’s apparent predictive power essentially evaporated.

The contrast with less rigorous evaluation strategies is striking and instructive. When the univariable filtering step was performed just once on the entire training cohort rather than repeated within each fold, the model’s estimated AUC rose to 0.637, with an optimism correction of 0.148 revealing how much the apparent performance of 0.785 had been inflated. On the small independent test set of 38 patients, the model posted an AUC of 0.716, a figure that looks respectable on paper but rests on so few cases that the confidence interval around it is wide enough to drive a truck through. Calibration analysis, which assesses whether predicted probabilities match observed outcomes, yielded a Brier score of 0.216, with a calibration slope of 1.502 suggesting predictions that were somewhat too timid in their extremes.

The team also explored, explicitly labeled as exploratory, a family of support vector machine models trained on three different feature sets: clinical factors alone, radiomic features alone, and a combination of both. Support vector machines are algorithms that seek the optimal boundary separating two classes in a high-dimensional space, and they remain popular in radiomics research for their ability to handle complex, nonlinear relationships. On the test set, the clinical-only model achieved an AUC of 0.603, the radiomics-only model 0.732, and the multimodal model 0.745. At first glance, the multimodal approach appears to win. But statistical testing told a different story. Using bootstrap paired comparisons with 5,000 resamples, the researchers found that the multimodal model’s advantage over the clinical-only model, a difference of 0.142 in AUC, carried a 95 percent confidence interval stretching from minus 0.077 to plus 0.364, with a p-value of 0.212. Against the radiomics-only model, the increment was a negligible 0.012, clearly non-significant.

This failure to demonstrate incremental value cuts to the core of the multimodal radiomics hypothesis. The entire rationale for combining imaging data with clinical variables is that each source contributes unique information. If the combined model performs no better than either component alone, the added complexity, computational cost, and validation burden of multimodal pipelines may not be justified. The study also examined whether age modified performance, a relevant question given the elderly cohort, and found multimodal discrimination was broadly similar across age strata, with AUCs of 0.739 versus 0.786 in the compared groups. Age, in other words, did not rescue the model’s performance or meaningfully stratify it.

Why did the radiomic signals prove so fragile? Several explanations deserve consideration. First, the sample size is modest by machine learning standards, and the 63 non-responders imposed a hard ceiling on model complexity. Second, radiomic features are notoriously sensitive to variations in CT acquisition parameters, reconstruction kernels, and contrast timing, all of which can vary even within a single center over time. Third, and perhaps most fundamentally, the feature selection process itself is a source of instability. When hundreds of candidate features are filtered and winnowed on slightly different subsets of data, different features can emerge as winners each time, a phenomenon the researchers probed with a 200-resample feature-stability analysis. The gap between the partially nested estimate of 0.637 and the fully nested estimate of 0.554 quantifies precisely how much of the model’s apparent skill was an artifact of information leakage rather than genuine biology.

The authors are refreshingly candid about these limitations, concluding that their pre-specified model achieves limited and unstable discrimination and that no evidence supports incremental value for combining clinical and radiomic data. They call for prospective multicenter validation before any clinical application, a bar that most published radiomics studies never clear. The study was conducted without external funding, supported entirely by the routine clinical work of the Department of Radiation Oncology, and approved by the ethics committee of Anhui Provincial Cancer Hospital with informed consent waived for the retrospective design.

In an era when headlines routinely proclaim that artificial intelligence can outperform physicians, this study offers a quieter and more valuable contribution: a demonstration of what rigorous, pre-specified, leakage-free evaluation actually looks like, and a reminder that the gap between apparent and real predictive performance can swallow an entire research program. For elderly patients with esophageal cancer, the promise of predicting treatment response from a routine CT scan remains just that, a promise. But by showing exactly where the promise breaks down, this proof-of-concept study may ultimately accelerate the field toward models that survive contact with honest statistics. The lesson for the radiomics community is clear: nested validation is not bureaucratic box-ticking. It is the difference between a model that works and a model that merely looks like it does.

Subject of Research: Prediction of early radiological response to radiotherapy in esophageal cancer using CT radiomics and clinical factors

Article Title: Prediction of early radiological response in esophageal cancer using CT radiomics and clinical factors: a single-center proof-of-concept study in an elderly cohort

Article References: Zhou, Q., Tao, Z.-C., Zhang, H.-B., Zhou, L.-R., & Liu, H.-W. (2026). Prediction of early radiological response in esophageal cancer using CT radiomics and clinical factors: a single-center proof-of-concept study in an elderly cohort. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02903-1

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02903-1

Keywords: esophageal cancer, radiomics, machine learning, CT imaging, radiotherapy, treatment response, LASSO logistic regression, support vector machine, nested cross-validation, predictive modeling, medical imaging, elderly patients

The clinical logic behind the study is straightforward but consequential. Small cell lung cancer is exquisitely radiosensitive, so outside true emergencies patients with this subtype may derive limited benefit from surgical resection of brain metastases. Large tumors from non-small cell lung cancer, by contrast, may warrant surgical removal before radiotherapy to relieve mass effect and improve local control. And a subset of non-small cell lung cancers harboring EGFR driver mutations may allow clinicians to defer upfront radiation entirely while trializing a tyrosine kinase inhibitor with intracranial activity. Three subtypes, three fundamentally different treatment pathways, and a diagnostic pipeline too slow to inform the moment when those pathways are chosen. The researchers therefore set out to build a tool that could estimate subtype using data available on day one.

The study team retrospectively identified 305 patients with metastatic lung cancer diagnosed between 2016 and 2025 at a single institution. In an unusual and methodologically deliberate design choice, the 182 patients who did not have brain metastases at diagnosis formed the training cohort, while the entire 123-patient group presenting with synchronous brain metastases was reserved as a held-out testing cohort. This partitioning preserved the clinically relevant target population for honest evaluation, though it introduced the possibility of distribution shift between the two groups, a concern the investigators addressed with dedicated sensitivity analyses. Patients were classified into three groups: small cell lung cancer, EGFR-mutated NSCLC, and EGFR-wild-type NSCLC. EGFR was singled out for independent categorization because it is among the most prevalent driver mutations and, unlike KRAS, has multiple central-nervous-system-penetrating targeted therapies already approved.

Clinical variables included sex, age, race, smoking history, pack-years, and prior lung disease such as emphysema or interstitial lung disease. The imaging features were extracted from narrative diagnostic chest CT reports using a firewalled institutional large language model, Anthropic Claude Sonnet 4.6, blinded to patient identifiers and tumor subtype. The model determined the presence or absence of fifteen predefined features, including tumor location, pleural attachment, cavitation, lymphangitic spread, fibrosis, miliary pattern, emphysema, mediastinal lymphadenopathy, vessel encasement, superior vena cava involvement, and osseous metastases. To validate this extraction, investigators manually reviewed fifty randomly selected reports as a reference standard, finding 96.8 percent overall agreement and a Cohen’s kappa of 0.918 across 750 individual feature comparisons. Notably, the authors emphasize that the modeling approach itself does not depend on the use of a large language model; the tool simply enabled consistent, high-throughput data extraction.

Two machine-learning approaches were tested: LASSO-regularized logistic regression, which performs feature selection while guarding against overfitting, and random forest, an ensemble method that aggregates many decision trees. Each was trained on clinical variables alone and on the combined clinical-plus-imaging feature set, and performance was evaluated with one-versus-rest AUC analysis in the held-out brain metastasis cohort. The combined LASSO logistic regression model achieved testing AUCs of 0.887 for EGFR-mutated disease, 0.811 for small cell lung cancer, and 0.801 for EGFR-wild-type disease. The random forest reached 0.907, 0.783, and 0.810 respectively. Adding imaging features significantly improved discrimination for EGFR-mutated cancer and small cell histology under logistic regression, and for small cell histology under random forest, while clinical variables alone were largely sufficient for the EGFR-wild-type group. Neither modeling approach demonstrated clear superiority over the other.

The three-way classification analysis offered perhaps the most clinically interpretable results. Both multiclass models showed their highest positive predictive value for EGFR-mutated cancer, exceeding 80 percent, with negative predictive values approaching 90 percent. For small cell lung cancer, the pattern inverted: positive predictive value was low, around 33 to 36 percent, but negative predictive value remained high, above 89 percent. In practical terms, the models are better at ruling out small cell histology than at confirming it, and patients predicted to have EGFR-mutated disease may represent a particularly informative group for early multidisciplinary discussion. An exploratory analysis of misclassified EGFR-wild-type patients revealed that several harbored other driver mutations, including KRAS G12A, ERBB2 mutations, and EML4-ALK fusions, suggesting the EGFR-wild-type category may partly capture a broader oncogenic-driver phenotype rather than a homogeneous entity.

Decision curve analysis added a layer of clinical realism by quantifying net benefit across a range of threshold probabilities, representing the minimum confidence at which a clinician would act on a predicted subtype. For EGFR-mutated cancer, the multiclass random forest outperformed both the flag-all and flag-none reference strategies across the entire examined threshold range from 0.05 to 0.80. At a representative threshold of 0.30, the models corresponded to roughly 12 to 14 additional correct EGFR-positive identifications per 100 patients evaluated without an increase in false positives. For small cell lung cancer, benefit was confined to low thresholds below approximately 0.30 and turned negative at higher thresholds, a limitation the authors attribute largely to the small number of small cell cases, just 19 of 123 testing patients, which constrains attainable positive predictive value and widens confidence intervals.

The feature-level analyses revealed biologically plausible signals. Miliary pulmonary metastases and pleural attachment were associated with EGFR-mutated disease, consistent with previously described imaging phenotypes of this molecular subtype, while smoking history, greater pack-years, superior vena cava involvement, and central thoracic disease distribution were associated with small cell histology. Cavitation and pleural attachment leaned toward EGFR-wild-type tumors. The investigators also confronted their design’s weaknesses directly: the single-institution cohort may overrepresent EGFR-mutated disease, imaging features were drawn from unstructured radiology reports rather than standardized interpretation, and the EGFR-wild-type group is biologically heterogeneous. Covariate reweighting and size-matched sensitivity analyses showed minimal change in model discrimination, but the authors stress that these checks cannot prove the training and testing populations were equivalent.

The bottom line is carefully bounded. The researchers explicitly state that these models should not substitute for histologic or molecular diagnosis, nor serve as a stand-alone basis for treatment decisions, and that generalizability requires validation in independent external cohorts. Yet as a proof of concept, the study demonstrates something genuinely useful: clinical and radiographic information already available at initial presentation can carry meaningful probabilistic information about lung cancer subtype during the diagnostic limbo that follows a devastating synchronous brain metastasis diagnosis. In that window, when neurosurgeons, radiation oncologists, and medical oncologists must weigh surgery, radiosurgery, and targeted therapy against one another, even an imperfect early estimate could sharpen multidisciplinary conversations and help prioritize which patients need expedited molecular confirmation. Prospective implementation studies will determine whether that promise translates into better outcomes, but the study makes a compelling case that the answers clinicians need may already be hiding in the charts and scans they see on day one.

Subject of Research: Machine-learning prediction of lung cancer subtype from clinical and chest CT features to guide early management of synchronous brain metastases

Article Title: Predictive modeling of lung cancer subtype with integration of clinical and thoracic imaging features to guide early brain metastasis management

Article References: Park, J., Buclez, P., Lubisich, J., Hanubal, K., Inda, A., Hui, C., Harris, J., & Simon, A. (2026). Predictive modeling of lung cancer subtype with integration of clinical and thoracic imaging features to guide early brain metastasis management. Journal of Neuro-Oncology, 180(1), Article 5. https://doi.org/10.1007/s11060-026-05812-z

Image Credits: AI Generated

DOI: 10.1007/s11060-026-05812-z

Keywords: lung cancer, brain metastases, EGFR mutation, small cell lung cancer, non-small cell lung cancer, machine learning, chest CT, predictive modeling, LASSO logistic regression, random forest, radiology reports, Journal of Neuro-Oncology