Construction and validation of a personalized nomogram of ultrasound for pretreatment prediction of breast cancer patients sensitive to neoadjuvant chemotherapy

Abstract

Objective: To construct a combined radiomics model based on pre-treatment ultrasound for predicting of advanced breast cancers sensitive to neoadjuvant chemotherapy (NAC).

Methods: A total of 288 eligible breast cancer patients who underwent NAC before surgery were enrolled in the retrospective study cohort. Radiomics features reflecting the phenotype of the pre-NAC tumors were extracted. With features selected using the least absolute shrinkage and selection operator (LASSO) regression, radiomics signature (Rad-score) was established based on the pre-NAC ultrasound. Then, radiomics nomogram of ultrasound (RU) was established on the basis of the best radiomic signature incorporating independent clinical features. The performance of RU was evaluated in terms of calibration curve, area under the curve (AUC), and decision curve analysis (DCA).

Results: Nine features were selected to construct the radiomics signature in the training cohort. Combined with independent clinical characteristics, the performance of RU for identifying Grade 4-5 patients was significantly superior than the clinical model and Rad-score alone (p < 0.05, as per the Delong test), which achieved an AUC of 0.863 (95% CI, 0.814-0.963) in the training group and 0.854 (95% CI, 0.776-0.931) in the validation group. DCA showed that this model satisfactory clinical utility, suggesting its robustness as a response predictor.

Conclusion: This study demonstrated that RU has a potential role in predicting drug-sensitive breast cancers.

Advances in knowledge: Aiming at early detection of Grade 4-5 breast cancer patients, the radiomics nomogram based on ultrasound has been approved as a promising indicator with high clinical utility. It is the first application of ultrasound-based radiomics nomogram to distinguish drug-sensitive breast cancers.

Figure 1.

Flowchart of this study with the exclusion criteria. NAC, neoadjuvant chemotherapy.

Figure 2.

(a, b) Workflow of the radiomics in this study. ROI, region of interest.

Figure 3.

Radiomics features selection with the LASSO. (a) Selection of the tuning parameter (λ) was conducted in the LASSO model via 10-fold cross-validation based on minimum criteria. The dotted vertical black line was delineated at the optimal value with the one standard error of the minimum and minimum criteria. The optimal λ value of 0.168 with log(λ) = −4.086 was selected. (b) LASSO coefficient profiles of the 464 radiomics features. The dotted vertical line was drawn at the optimal value selected by 10-fold cross-validation, where optimal result was nine non-zero coefficient. LASSO, least absolute shrinkage and selection operator

Figure 4.

Nomogram to predict G4-5 patients before NAC by combining clinical characteristic and Rad-score. NAC, neoadjuvant chemotherapy.

Figure 5.

(a, b) Comparison of ROCs of the clinical model, RS model, and the RU model in the training and validation sets. (c, d) Calibration curve of the RU model in the training cohort and testing cohort. The X-axis is on behalf the predictive probability of Grade 4–5 cases; the Y-axis represents the actual rate of Grade 4–5 patients. The blue line denotes the ideal prediction of Grade 4–5 patients, and the pink line demonstrates the performance of the RU model of which the closer the red line accords with the perfect line represents the better prediction. ROC, receiver operating characteristic.

Figure 6.

Decision curve of the RS model, clinical model, and the RU model. The black line manifests the hypothesis that all cases are Grade 1–3 breast cancer; the gray line denotes the hypotheses that all patients achieved Grade 4–5. The y-axis represents the net benefit, and the x-axis manifests the high-risk threshold. The blue, orange, and pink lines indicate the RS model, clinical model, and RU model.