HER2 as a Therapeutic Target. Bladder cancer is the third most prevalent cancer type exhibiting human epidermal growth factor receptor 2 (HER2) overexpression. HER2 plays a key role in driving tumor growth and is associated with pathological malignancy and poor prognosis. Accurate determination of HER2 status directly determines treatment eligibility for HER2-targeted therapies, particularly antibody-drug conjugates (ADCs), which have shown strong efficacy in HER2-positive bladder cancer.
FDA-Approved HER2-Targeted Therapy. A multicenter phase II clinical trial involving 107 patients demonstrated that disitamab vedotin, a novel ADC targeting HER2, produced an objective response rate of 50.5% in patients with HER2-positive locally advanced or metastatic urothelial carcinoma who had received at least one prior line of chemotherapy. The FDA has since approved disitamab vedotin for this indication, making accurate pre-treatment HER2 status assessment clinically urgent.
Limitations of Current HER2 Testing. HER2 status is currently assessed by immunohistochemistry (IHC) staining of transurethral cystoscopy biopsy or surgical specimens. However, intra-tumor heterogeneity and sampling limitations mean that partial specimens may not represent the entire tumor. IHC is also invasive, time-consuming, and limited in reproducibility, making non-invasive alternatives desirable, particularly for patients who cannot safely undergo repeated biopsies.
Radiomics as a Non-Invasive Alternative. Radiomics extracts large numbers of quantitative features from CT and MRI images that reflect tumor heterogeneity and biological behavior. Prior studies have shown that radiomics can assess benign-versus-malignant status, tumor grading, staging, and recurrence in bladder cancer, and has successfully predicted HER2, PD-L1, and Ki-67 expression in breast, gastric, and liver cancers from preoperative imaging. This study is the first multicenter CT-based radiomics study to predict HER2 status specifically in bladder cancer.
Patient Population. 207 patients with pathologically confirmed bladder cancer who underwent radical or partial cystectomy were retrospectively recruited from four independent hospitals. The training set comprised 154 patients from the primary center (June 2015 to June 2023), and the test set comprised 53 patients from three other centers (June 2019 to December 2023). Patients who received neoadjuvant therapy, had non-urothelial carcinoma, or had incomplete CT data were excluded.
HER2 Status Assessment. HER2 expression was determined by IHC staining of formalin-fixed, paraffin-embedded surgical specimens, scored by two pathologists blinded to clinical data according to the 2018 American Society of Clinical Oncology guidelines. Scores of 0 (no staining) and 1+ (weak partial membrane) were classified as HER2-negative; scores of 2+ (weak to moderate complete membrane) and 3+ (strong complete membrane) were classified as HER2-positive. Overall, 109 patients (52.7%) were HER2-positive and 98 (47.3%) were HER2-negative.
CT Image Acquisition and ROI Segmentation. All patients underwent contrast-enhanced CT within 2 weeks before surgery. Nephrographic phase CT images -- the most informative phase for bladder tumor identification -- were selected for analysis. A radiologist with 5 years of urological imaging experience manually segmented the tumor region of interest in ITK-SNAP software, blinded to pathological data. For patients with multiple tumors, the largest lesion was selected. Inter- and intra-observer reproducibility was assessed using intraclass correlation coefficients (ICC), with ICC greater than 0.75 required for feature retention.
Feature Extraction Pipeline. 1,218 radiomics features were extracted from the 3D regions of interest using the open-source PyRadiomics package. Features were standardized using z-score normalization and then progressively filtered: 887 features passed the ICC reliability threshold; 223 remained after removing highly correlated features; 33 remained after t-test comparison between HER2-positive and HER2-negative groups; and finally LASSO regression with fivefold cross-validation selected 11 features for model construction.
Five ML Algorithms Compared. The 11 selected radiomics features were used to train and compare five machine learning classifiers: logistic regression (LR), random forest (RF), support vector machine (SVM), extreme gradient boosting (XGBoost), and k-nearest neighbors (KNN). All models were trained on the primary center cohort and evaluated on the independent three-center test set using area under the ROC curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
Random Forest as the Selected Model. Random forest is an ensemble learning method that builds multiple decision trees during training and outputs the class predicted by the majority of trees. Its strength lies in handling high-dimensional feature spaces with correlated variables while providing natural resistance to overfitting through tree averaging. The model's internal structure can be visualized by examining individual decision trees, providing an additional layer of interpretability beyond statistical metrics.
SHAP for Model Interpretability. The inherent black-box nature of machine learning models limits clinical trust and adoption. SHAP (Shapley Additive Explanations), derived from game theory, quantifies each feature's individual contribution to each prediction by computing its marginal contribution across all possible feature combinations. Global SHAP analysis identifies which features most consistently drive model predictions across all patients, while local SHAP force plots explain why the model made its prediction for each individual patient, enabling case-by-case clinical reasoning.
Statistical Analysis Framework. Model discrimination was assessed by AUC values with 95% confidence intervals calculated by bootstrapping. Delong's test was used to statistically compare AUC values between models. Calibration curves and performance radar plots were generated for all models in both training and test sets. Clinical and pathological characteristics between training and test sets were compared using chi-square test, Fisher exact test, t-test, and Mann-Whitney U test as appropriate.
Training Set Performance. In the training set, random forest achieved the highest AUC of 0.965 (95% CI: 0.933-0.987), followed closely by XGBoost at 0.961, SVM at 0.917, LR at 0.829, and KNN at 0.827. The RF model's training set accuracy was 0.916, sensitivity 0.944, and specificity 0.892. Delong's test confirmed that RF achieved significantly higher AUC than LR (p less than 0.001), SVM (p = 0.035), and KNN (p less than 0.001), but not XGBoost (p = 0.643).
Independent Test Set Performance. In the independent multicenter test set, RF again achieved the highest AUC of 0.815 (95% CI: 0.695-0.917), outperforming LR (AUC 0.803), XGBoost (AUC 0.794), SVM (AUC 0.709), and KNN (AUC 0.679). The RF model achieved test set accuracy of 0.755, sensitivity of 0.667, specificity of 0.846, PPV of 0.710, and NPV of 0.818. Despite the lack of significant AUC difference from XGBoost in training, RF achieved higher accuracy, sensitivity, PPV, and NPV in both training and test sets, making it the preferred final model.
Consistent Performance Across Training and Test Sets. The RF model maintained performance from training (AUC 0.965) to multicenter test (AUC 0.815), representing acceptable generalization despite the training-to-test AUC gap that is typical in radiomics studies. This gap reflects differences in the distribution of pathological T-stage and N-stage between the training and test cohorts -- the test set contained a higher proportion of T2 and N1-2 disease -- introducing some distributional shift that would challenge any model.
Comparison to Prior Radiomics Models. The RF model's test AUC of 0.815 for HER2 prediction in bladder cancer is comparable to radiomics-based HER2 prediction models in breast and gastric cancer. Prior CT-based radiomics models predicting PD-L1 and Ki-67 expression in bladder cancer achieved validation AUC values of 0.753-0.887, positioning this HER2 prediction model within the established performance range for CT radiomics biomarker prediction in bladder cancer.
Top Predictive Features Identified. SHAP global analysis revealed that the two highest-importance features in the RF model were log-sigma-1-0-mm-3D_glszm_SmallAreaEmphasis and log-sigma-2-0-mm-3D_glszm_SmallAreaEmphasis -- both gray level size zone matrix (GLSZM) features applied to Laplacian of Gaussian filtered images at different sigma values. The third most important feature was a wavelet-transformed gray level co-occurrence matrix (GLCM) feature. These texture features capture fine-grained spatial heterogeneity within the tumor that correlates with HER2 biological status.
What GLSZM SmallAreaEmphasis Measures. Gray level size zone matrix analysis characterizes runs of pixels with similar intensity in multiple directions. The SmallAreaEmphasis feature specifically quantifies the distribution of small-scale zones within the texture, measuring the fineness of tumor texture at different spatial scales captured by the Laplacian of Gaussian filter. Tumors with higher SmallAreaEmphasis values have more fine-grained texture heterogeneity. A prior study in gastric cancer identified SmallAreaEmphasis as an important predictor of HER2 expression, suggesting a consistent biological mechanism linking fine texture heterogeneity with HER2 overexpression across cancer types.
Local SHAP Force Plots for Individual Cases. In addition to global feature importance, the study generated SHAP force plots for four representative correctly-predicted cases (two HER2-positive, two HER2-negative), visualizing how individual feature values pushed the prediction toward positive or negative status for each patient. This local interpretability allows clinicians to understand why the model classified a specific patient as HER2-positive, identifying which imaging features most contributed to that prediction -- directly addressing the clinical trust barrier that limits machine learning adoption.
Biological Rationale for Texture-HER2 Links. HER2 overexpression drives tumor cell proliferation, architectural disorganization, and microenvironmental heterogeneity. These biological processes manifest as changes in CT image texture that GLSZM and GLCM features are designed to capture. The multi-scale analysis using different Laplacian of Gaussian sigma values captures heterogeneity at both fine and coarse spatial scales, aligning with the known multiscale spatial disorganization associated with high HER2 expression in bladder cancer tissue.
Non-Invasive HER2 Assessment as Clinical Tool. This study provides the first multicenter CT radiomics model for non-invasive preoperative HER2 status prediction in bladder cancer. A model achieving AUC of 0.815 and specificity of 0.846 with high NPV of 0.818 on a multicenter test set could serve as a non-invasive screening tool to identify HER2-negative patients with high confidence, potentially reducing the need for invasive IHC testing in patients who cannot safely undergo biopsy and enabling earlier treatment planning in patients with confirmed HER2-positive disease.
Advantages Over Existing Testing Approaches. Unlike IHC testing, CT radiomics analysis can capture the entire tumor volume rather than a single biopsy specimen, potentially overcoming the intra-tumor heterogeneity limitation that causes focal sampling to misclassify HER2 status. CT is already routinely acquired as part of pre-surgical bladder cancer workup, meaning the radiomics analysis adds no additional radiation exposure or cost beyond what patients already undergo. The multicenter validation across four institutions supports generalizability across different CT scanners and scanning protocols.
Study Limitations. The retrospective design with relatively small sample size across four centers limits statistical power and introduces potential selection bias. Differences in pathological T-stage and N-stage distribution between training and test cohorts may have affected test set performance. While SHAP analysis identified the most important radiomics features, the biological connections between specific CT texture features and HER2 genomic or protein expression mechanisms require further investigation to fully validate the biological rationale for these image-to-biology associations.
Path to Automated and Prospective Validation. Manual ROI segmentation by radiologists is time-consuming and introduces inter-observer variability that limits clinical scalability. Future work should focus on automated deep learning-based segmentation combined with end-to-end radiomics pipelines to eliminate human intervention and improve reproducibility. Prospective multicenter validation studies with larger patient cohorts, standardized CT acquisition protocols, and integration with clinical decision support systems are needed before this model can be recommended for routine clinical use to guide HER2-targeted therapy selection.