Noninvasive Artificial Intelligence System for Early Predicting Residual Cancer Burden During Neoadjuvant Chemotherapy in Breast Cancer

Abstract

Objective: To develop an artificial intelligence (AI) system for the early prediction of residual cancer burden (RCB) scores during neoadjuvant chemotherapy (NAC) in breast cancer.

Background: RCB III indicates drug resistance in breast cancer, and early detection methods are lacking.

Methods: This study enrolled 1048 patients with breast cancer from 4 institutions, who were all receiving NAC. Magnetic resonance images were collected at the pre-NAC and mid-NAC stages, and radiomics and deep learning features were extracted. A multitask AI system was developed to classify patients into 3 groups (RCB 0 to I, II, and III ) in the primary cohort (PC, n=335). Feature selection was conducted using the Mann-Whitney U test, Spearman analysis, least absolute shrinkage and selection operator regression, and the Boruta algorithm. Single-modality models were developed, followed by model integration. The AI system was validated in 3 external validation cohorts (EVCs, n=713).

Results: Among the patients, 442 (42.18%) were RCB 0 to I, 462 (44.08%) were RCB II, and 144 (13.74%) were RCB III. Model I achieved an area under the curve of 0.975 in the PC and 0.923 in the EVCs for differentiating RCB III from RCB 0 to II. Model II distinguished RCB 0 to I from RCB II-III, with an area under the curve of 0.976 in the PC and 0.910 in the EVCs. Subgroup analysis confirmed that the AI system was consistent across different clinical T stages and molecular subtypes.

Conclusions: The multitask AI system offers a noninvasive tool for the early prediction of RCB scores in breast cancer, supporting clinical decision-making during NAC.

FIGURE 1

Flowchart of this study. Pre-NAC and mid-NAC multiparametric MRI and baseline data were collected. The data from the primary cohort were used for model development, and the data from the other 3 cohorts were used as independent validation cohorts. This study included radiomic and deep learning feature engineering, feature selection, model development, model evaluation, and clinical application assessment.

FIGURE 2

(A–D) ROC curves among different cohorts for distinguishing the RCB III and RCB 0 to II; (D–H) ROC curves among different cohorts for distinguishing the RCB 0 to I and RCB II to III.

FIGURE 3

Representative MRI and tumor delineation in pre-NAC and mid-NAC stages from 2 randomly selected patients. (A and B) The whole tumor region on MRI before and at the mid-stage of NAC treatment, showing that the tumor size decreased; (C) The final pathology report showed that the patient achieved RCB 0; (D and E) The whole tumor region on MRI before and at the mid-stage of NAC treatment, showing that the tumor size did not decrease, although the shape changed; (F) The final pathology report showed that the patient achieved RCB III; (G–J) The confusion matrix of true RCB groups and model prediction results among different cohorts.

FIGURE 4

Histograms of LRadDP for predicting RCB scores in different groups based on subgroup analysis, including clinical T stage (A), molecular subtype (B) and treatment period (C). The data are presented as percentages. True positives indicate that the patient was predicted to be in the true group, and false positives indicate that the patient was predicted to be in the false group.