Development and validation of a radiopathomic model for predicting pathologic complete response to neoadjuvant chemotherapy in breast cancer patients

Abstract

Background: Neoadjuvant chemotherapy (NAC) has become the standard therapeutic option for early high-risk and locally advanced breast cancer. However, response rates to NAC vary between patients, causing delays in treatment and affecting the prognosis for patients who do not sensitive to NAC.

Materials and methods: In total, 211 breast cancer patients who completed NAC (training set: 155, validation set: 56) were retrospectively enrolled. we developed a deep learning radiopathomics model(DLRPM) by Support Vector Machine (SVM) method based on clinicopathological features, radiomics features, and pathomics features. Furthermore, we comprehensively validated the DLRPM and compared it with three single-scale signatures.

Results: DLRPM had favourable performance for the prediction of pathological complete response (pCR) in the training set (AUC 0.933[95% CI 0.895-0.971]), and in the validation set (AUC 0.927 [95% CI 0.858-0.996]). In the validation set, DLRPM also significantly outperformed the radiomics signature (AUC 0.821[0.700-0.942]), pathomics signature (AUC 0.766[0.629-0.903]), and deep learning pathomics signature (AUC 0.804[0.683-0.925]) (all p < 0.05). The calibration curves and decision curve analysis also indicated the clinical effectiveness of the DLRPM.

Conclusions: DLRPM can help clinicians accurately predict the efficacy of NAC before treatment, highlighting the potential of artificial intelligence to improve the personalized treatment of breast cancer patients.

Keywords: Breast cancer; CECT; Pathomics; Radiomics; Radiopathomics.

Fig. 1

Flow diagram of patient cohort selection

Fig. 2

Workflow of Study. The images were preprocessed for feature extraction. After feature evaluation and model construction, four sets of features [radiomic signature (RS), pathologic signature (PS), deep learning pathologic signature (DLPS) and clinical features] were generated and further used to construct DLRPM. The performance of DLRPM in predicting pCR before NAC was validated in validation set

Fig. 3

CECT and histology images from complete responder (A) and partial responder (B) before NAC and after 8 courses of NAC. In the CECT image, it is seen that the tumor in the complete responder have completely dissipated in the post-NAC image, and stromal tissue with no visible tumor cells was presented in the pathological images. But CECT and histology images from and partial responder shows residual tumor cells but reduced compared to baseline

Fig. 4

Feature selection process. Radiomics features (A, B) and pathomics features (C, D) were selected by the LASSO model with tuning parameter (λ) using fivefold cross-validation via minimum and 1se criteria

Fig. 5

The raincloud plot visualizes prediction probability of RS, PS and DLPS, it shows the sample distribution locations and interval sample densities for the training (A)and validation set(B) of signatures

Fig. 6

ROC analysis of predict models for predicting pCR in the training set (A) and validation set (B), respectively. C Calibration curves of models in training set on discriminating Non-pCR versus pCR. D Decision curve analysis in training set using RS, PS, DLPS and DLRPM