CT Radiomics Adds Nothing Over Tumor Volume for Early Immunotherapy Response

For patients with metastatic melanoma, immune checkpoint inhibitors can produce durable, sometimes years-long remissions, but a substantial share of patients gain no benefit from these drugs. Clinicians therefore want early signals that reveal whether a treatment is working before precious months are lost. A new retrospective study now delivers a sobering answer to one proposed solution: sophisticated CT radiomics, the quantitative analysis of tumor texture and intensity, offered no conclusive improvement over simple measurements of tumor volume for predicting how individual metastatic lesions would respond early in therapy.

The study, published in the journal Cancer Imaging, was conducted by a multicenter team led by Seval Akbal and Lukas T. Rotkopf. The researchers evaluated 158 patients with 1,626 evaluable metastatic lesions drawn from two independent cohorts: an internal cohort of 129 patients imaged with standardized CT protocols at a single academic cancer center between May 2013 and October 2023, and an external cohort of 29 patients scanned at a collaborating institution between March 2021 and July 2024 using heterogeneous scanners and acquisition parameters that reflect everyday clinical practice. The primary endpoint was binary lesion-level size response at the second follow-up scan, distinguishing clinical benefit, meaning complete response, partial response, or stable disease, from progressive disease.

The design was deliberately structured to test generalizability. Models were trained on the internal cohort using four-fold patient-grouped cross-validation and then applied once, unchanged and without any center indicator, to the external cohort. Five nested Random Forest predictor sets were compared: clinical variables alone, volumetric features alone, radiomics alone, volumetric plus radiomics, and a comprehensive combination. The volumetric set consisted simply of log-transformed lesion volumes at baseline and first follow-up plus the interval between scans. Radiomic features, 72 per lesion per timepoint, were extracted in compliance with the Image Biomarker Standardization Initiative and included first-order histogram statistics, intensity-based features, and gray-level co-occurrence matrix texture features, with delta features capturing changes from baseline to first follow-up. Clinical features included age, sex, treatment regimen, time since diagnosis, number of affected organs, and baseline lactate dehydrogenase.

The headline result was clear. For predicting response at the second follow-up, the volumetric model achieved an internal area under the receiver operating characteristic curve of 0.84, with a 95 percent confidence interval of 0.78 to 0.88, while the radiomics model reached 0.86, with a confidence interval of 0.81 to 0.90. When the frozen models were applied to the external cohort, which contained only 12 progression events, both the volumetric and radiomics models achieved an AUC of 0.68, with confidence intervals spanning roughly 0.50 to 0.88. No significant difference emerged between any of the models in either cohort. The researchers formally tested equivalence using paired differences in AUC against a margin of plus or minus 0.05, concluding equivalence only when the entire confidence interval of the difference fell within that margin.

Baseline-only prediction told an even starker story. When only features from the initial scan were available to predict response at the first follow-up, all models performed near chance, and none exceeded the volumetric reference. Discrimination improved substantially only once early lesion dynamics from the first follow-up entered the models, and it was the volume dynamics themselves that captured the majority of the predictive signal. Notably, early volume change alone, without any fitted model at all, retained most of the volumetric model’s discrimination internally and even exceeded it externally.

A central motivation for the study was the possibility that radiomic features might merely be indirect proxies for tumor size. Many texture and histogram-based features correlate strongly with lesion volume, meaning radiomics models could be capturing volume dynamics through the back door rather than detecting independent biological characteristics. To probe this, the team computed Spearman correlations between lesion volume and all 288 radiomic features. The median absolute correlation was 0.37, but a minority of features were strongly volume-dependent: 19 features, or 6.6 percent, showed correlations above 0.7 in absolute value, and 8 features, or 2.8 percent, exceeded 0.9, with the strongest reaching 0.98. Importantly, volume-residualized, volume-decorrelated, and redundancy-filtered radiomics retained their discrimination yet still failed to improve on volume dynamics, suggesting that the redundancy question is more nuanced than simple confounding.

The researchers also asked whether radiomics might help in specific anatomical settings, since lesions in the same patient often respond differently under checkpoint inhibition, a fact that matters for decisions about oligoprogression. In exploratory organ-specific subgroup analyses of the internal cohort, no lesion site showed a statistically significant improvement from adding radiomics to volumetric features, although all subgroups were underpowered. Across lesion-size subgroups, no group favored radiomics after adjustment for multiple comparisons. The largest lesions, in which texture can be calculated most stably, showed a numerical gain that did not survive correction, so the authors caution that noise-limited texture measurements in small lesions cannot be excluded as a contributor to the negative result.

Sensitivity analyses reinforced the main conclusion. Alternative learners and repeated cross-validation did not change the comparison. ComBat harmonization, a statistical technique for reducing batch effects, did not improve the external discrimination of the radiomics model. Cross-center distribution filtering produced a numerically higher external radiomics AUC than volume dynamics, but without a significant difference, and an internal gain from the distribution-filtered comprehensive model that was significant after Holm adjustment failed to replicate externally. An extended feature set with higher-order texture families, spatial filters, and four bin widths did not exceed the volumetric reference. Inter-observer agreement from an independent second reading was high for first-order and texture features, arguing against contour noise as the sole explanation for the findings.

The results align with a growing body of skeptical literature. Previous work has found no significant advantage of adding radiomics to clinical baseline models for predicting early response in checkpoint inhibitor treatment, and a deep learning model trained directly on CT lesion images was likewise no better than a pure clinical model at forecasting treatment outcome. Earlier positive reports, the authors note, differ from this study in key respects: they were typically single-center efforts with standardized protocols, baseline-only signatures, patient-level endpoints, and small cohorts with limited validation. The negative result here concerns only incremental value over volume dynamics for a size-based lesion-level endpoint and does not invalidate associations reported for other endpoints such as survival.

The study has important limitations that temper how far the conclusion can be pushed. The external cohort was small, with only 29 patients contributing 12 progression events, so its confidence intervals were too wide to establish either equivalence or a deficit externally. Treatment distribution differed substantially between cohorts, with combination immunotherapy predominating externally, and because regimen was closely tied to center, treatment-related differences could not be separated from acquisition differences. The lesion-level analysis treats observations from the same patient as partially independent, which may inflate performance estimates, although patient-grouped cross-validation, cluster bootstrap resampling, and equal-patient weighting addressed this concern. Second follow-up results also include only surviving patients, a bias that inverse-probability weighting did not materially change. The authors scored their own study 18 of 36 points on the Radiomics Quality Score. Their conclusion is measured: volumetric features capture the primary predictive information for early lesion-level response, CT radiomics showed no conclusive improvement over volume dynamics, and confirmation in larger independent cohorts is required before radiomic signatures can be recommended for this task. Future work, they suggest, should explore whether deep learning approaches can extract predictive information from imaging beyond what volume dynamics already provide.

Subject of Research: CT radiomics versus tumor volume dynamics for predicting early lesion-level response to immunotherapy in metastatic melanoma

Article Title: CT radiomics showed no improvement beyond volume dynamics for early lesion-level size response to immunotherapy in metastatic melanoma

Article References: Akbal, S., Schepers, J. C., Op gen Oorth, P. V., Bleymehl, V., Schroth, M., Winneknecht, F., Machiraju, D., Holzschuh, J. C., Zhang, K. S., Wohlfeil, S., Ayx, I., Schoenberg, S. O., Schlemmer, H.-P., Hassel, J. C., & Rotkopf, L. T. (2026). CT radiomics showed no improvement beyond volume dynamics for early lesion-level size response to immunotherapy in metastatic melanoma. Cancer Imaging, 26(1), Article 138. https://doi.org/10.1186/s40644-026-01135-4

Image Credits: AI Generated

DOI: 10.1186/s40644-026-01135-4

Keywords: CT radiomics, metastatic melanoma, immune checkpoint inhibitors, tumor volume, lesion-level response, prediction models, AUC, iRECIST, biomarkers, multicenter study, texture features, external validation