Pathological Complete Response (pCR)

Updated 28 November 2025

pCR is a binary histopathologic endpoint defined by the complete absence of invasive carcinoma in the breast and lymph nodes post-neoadjuvant chemotherapy.
It is a key surrogate marker, correlating with improved survival outcomes, especially in aggressive subtypes like HER2+ and triple-negative breast cancers.
Advanced imaging, radiomics, and deep learning methods drive precise pCR prediction and support adaptive, personalized therapy strategies.

Pathological complete response (pCR) is a rigorously defined binary histopathologic endpoint signifying the total eradication of invasive carcinoma from both breast parenchyma and regional lymph nodes following neoadjuvant chemotherapy (NAC) in breast cancer. As a surrogate marker, pCR offers a reliable, early indication of therapy efficacy and is a central outcome in clinical trials and translational oncology research. Comprehensive understanding of pCR encompasses clinical definition, quantification, association with survival outcomes, predictive modeling frameworks, imaging and histopathological characterization, and the implications for adaptive therapy paradigms.

1. Definition and Clinical Role of Pathological Complete Response

pCR is universally defined as the absence of residual invasive cancer on histopathologic examination of both the breast tissue and axillary lymph nodes after completion of NAC, typically designated as ypT0/Tis ypN0 in TNM staging (Gilad et al., 2022, Fridman et al., 13 Jun 2025, Yang et al., 2023, Khan et al., 20 May 2025, Wang et al., 2019, Wang et al., 5 Nov 2024, Malhaire et al., 21 Nov 2025). In clinical trials and registries such as ACRIN-6698, I-SPY1, I-SPY2, and Duke, pCR labeling is directly abstracted from standardized clinical pathology reports.

pCR status stratifies patients as follows:

pCR (responder): No residual invasive carcinoma in breast and sampled lymph nodes; in situ lesions (e.g., DCIS) may persist.
non-pCR (non-responder): Any presence of residual invasive carcinoma.

Residual Cancer Burden (RCB) provides a finer-grained ordinal quantification, where RCB-0 equates to pCR and higher classes (RCB-I-III) represent increasing burden (Malhaire et al., 21 Nov 2025). The prevalence of pCR varies by cohort and tumor subtype, ranging from 19% in luminal tumors to >55% in HER2+ and triple-negative breast cancers (Malhaire et al., 21 Nov 2025, Khan et al., 20 May 2025).

Achieving pCR after NAC is strongly prognostic for improved event-free and overall survival, especially in aggressive subtypes (Gilad et al., 2022, Tai et al., 13 May 2024, Tan et al., 2021). This robust prognostic value underpins its use as a surrogate endpoint in drug registration trials and for risk-adaptive post-NAC management.

2. pCR as a Surrogate Endpoint and its Causal Interpretation

The clinical significance of pCR transcends simple correlation with survival outcomes. Principal stratification frameworks rigorously target the causal effect of treatment in the subpopulation who would attain pCR under intervention. Let $Z_i \in \{0,1\}$ denote randomization to control or novel therapy, $S_i(z)$ be potential pCR status under $z$ , and $Y_i(z)$ the long-term endpoint (e.g., survival).

The principal-stratum estimand,

$\theta = \mathbb{E}[Y_i(1)-Y_i(0)\mid S_i(1)=1],$

quantifies the treatment-induced survival difference among "potential responders" (Tan et al., 2021). Identification requires randomization, monotonicity ( $S_i(1) \ge S_i(0)$ ), and parametric modeling for latent subgroups. Empirical estimation strategies involve moment-matching and inverse-probability decompositions; for right-censored survival, Kaplan-Meier rates replace proportions. In NSABP B-40, the principal effect for $S(1)=1$ patients was a 16–18 percentage point improvement in 3-year event-free survival, confirming pCR's partial mediating role between therapy and outcome.

3. Imaging and Histopathologic Predictors of pCR

The technical landscape for pCR prediction is dominated by MRI–based radiomics, deep learning, and hybrid clinical-imaging models.

MRI parameters: DCE-MRI and DWI yield quantitative descriptors (kinetic, pharmacokinetic, ADC, synthetic CDI $^s$ ) routinely extracted from tumor ROIs (Kim et al., 5 Jun 2024, Tai et al., 13 May 2024, Gilad et al., 2022, Wang et al., 5 Nov 2024). Radiomics features (first-order statistics, shape, texture—GLCM, GLRLM, NGTDM) are engineered or learned via deep feature extraction (e.g., CRBM, ResNet-34, ViT) (Wang et al., 2019, Fridman et al., 13 Jun 2025).

Histopathologic correlates: Deep models applied to H&E-stained whole-slide images leverage multiple instance learning and attention mechanisms to infer pCR, with model attention frequently colocalizing to regions of heightened immune infiltrates (CD8+, PD-L1+, CD163+ cells), established predictors of chemo-sensitivity in triple-negative breast cancer (Khan et al., 20 May 2025).

Semantic MRI features: Morphologic patterns—oval/round shape, non-spiculated margins, unifocality, absence of non-mass enhancement, and smaller pre-NAC tumor size—are independently associated with higher pCR likelihood (Malhaire et al., 21 Nov 2025). BI-RADS–based descriptors (focality, margin type) and tumor biology (Ki67, TILs, subtype) confer additional discriminative power.

4. Predictive Modeling Approaches and Quantitative Results

Research converges on a suite of high-performance machine learning architectures, typically underpinned by imaging, radiomics, and clinical features.

DWI/DCE-MRI models: The "PD-DWI" XGBoost radiomics model, incorporating physiologically decomposed DWI features and clinical covariates, achieved a test AUC of 0.885 and $F_1$ of 0.739 on the ACRIN 6698/BMMR2 dataset (Gilad et al., 2022). Optimized synthetic CDI $^s$ fused with DWI and volumetric deep radiomics achieved 93.3% LOOCV accuracy and $F_1$ of 0.90 (Tai et al., 13 May 2024).

Transformer and attention models: Vision Transformer (ViT) models combining three DCE-MRI contrast phases attained AUC 0.94 and accuracy 0.93 in HR+/HER2– cohorts (BreastDCEDL), setting state-of-the-art performance benchmarks (Fridman et al., 13 Jun 2025). TopoTxR, a topology-guided 3D CNN, explicitly extracts persistent loops and voids representing fibroglandular architecture, yielding a 4.6 percentage point AUC improvement over dense CNNs (I-SPY1 AUC 0.917, accuracy 0.931) (Wang et al., 5 Nov 2024).

Histology-based networks: ECDEDL ensemble models, partitioning tiles into tumor and stroma subsets, showed external validation AUC gains of 6.2 points versus direct models (AUC 0.68), with accuracy increase from 56% to 71% (Yang et al., 2023).

Temporal and multimodal fusion: Temporal phenotype trajectories from four-point DCE-MRI boosted balanced accuracy for pCR from 0.76 (pre-NAC) to 0.86 (all time points) (Janíčková et al., 18 Sep 2025). Multimodal transformers integrating DCE-MRI and ADC radiomics realized AUC 0.76, significantly exceeding single-modality encoders (Kim et al., 5 Jun 2024).

Logistic/random forest models: Pretreatment MRI descriptors (shape, margins, size) added to clinicobiological covariates in random forest models increased pCR prediction sensitivity and precision (sens. 0.67, precision 0.71, AUC 0.66), compared to clinical variables alone (Malhaire et al., 21 Nov 2025).

Method	Modality	Test AUC	Accuracy	F1 Score
PD-DWI (XGB)	DWI, clinical	0.885	—	0.739
ViT-B/16	DCE-MRI (RGB fusion)	0.94*	0.93*	—
TopoTxR	DCE-MRI, topology	0.917	0.931	—
ECDEDL	Histopathology (WSI)	0.68	0.71	—
TemporalRep	DCE-MRI, 4-point traj.	0.86	—	—
Synthetic CDI $^s$	DWI, deep radiomics	—	0.93	0.90

*HR+/HER2– subgroup (Fridman et al., 13 Jun 2025)

Performance is subject to dataset composition, imaging protocol, and evaluation metric. Several studies report higher internal validation than external validation results, highlighting generalization challenges due to domain shift.

5. Datasets, Evaluation Metrics, and Validation Strategies

pCR studies leverage large, harmonized public datasets (I-SPY1, I-SPY2, ACRIN-6698, Duke, MAMA-MIA, BMMR2), with per-patient ground-truth encoded as binary pCR/non-pCR based on standardized pathology (Fridman et al., 13 Jun 2025, Tai et al., 13 May 2024, Musah, 3 Aug 2025). Data preprocessing includes strict anatomical cropping, voxel-level intensity normalization, and, where required, domain harmonization.

Evaluation follows established conventions:

Primary: AUROC (Area under the Receiver Operating Characteristic), balanced accuracy, $F_1$ score.
Secondary: sensitivity, specificity, precision, negative predictive value.

Cross-validation protocols (five- or ten-fold), nested validation, fixed train/validation/test splits, and LOOCV are used for unbiased measurement, with stratification for class balance (Fridman et al., 13 Jun 2025, Tai et al., 13 May 2024, Janíčková et al., 18 Sep 2025). Statistical comparisons are made via DeLong’s test (AUROC), McNemar’s test (accuracy), or paired $t$ -tests.

6. Current Limitations and Prospects for Clinical Translation

Despite high cross-validated performance in retrospective cohorts, limitations persist:

External validation AUCs often lag internal results, reflecting institutional staining, acquisition, and demographic variability (Yang et al., 2023).
Modest sensitivity of some models in stratified or rare subgroups (e.g., sensitivity 0.27 in ViT overall test set, but better in HR+/HER2–) (Fridman et al., 13 Jun 2025).
Class imbalance and fairness: pCR is less frequent than non-pCR in most cohorts, complicating sensitivity/recall and subgroup calibration (Musah, 3 Aug 2025).
Black-box models face challenges of explainability and clinical trust.

Future research priorities include federated and multicenter generalization, integration of clinical, imaging, and multi-omics data, interpretability tooling (heatmaps, attention), and real-time deployment in clinical workflows (Kim et al., 5 Jun 2024, Yang et al., 2023, Jing et al., 28 Jan 2025). Prospective clinical studies are necessary to solidify the clinical impact of pCR pre- and mid-treatment prediction as a tool for adaptive neoadjuvant therapy design.

7. Clinical Impact and Adaptive Therapy

Accurate early or pre-NAC pCR prediction enables:

Therapy adaptation: Early intensification or regimen changes for likely non-responders, potentially improving outcome (Jing et al., 28 Jan 2025, Janíčková et al., 18 Sep 2025).
De-escalation: Sparing likely responders from unnecessary toxicity while safely enabling breast-conserving approaches (Khan et al., 20 May 2025).
Trial design: Accelerated approval pathways for new agents using pCR as an intermediate endpoint, supported by principal-stratum causal estimands (Tan et al., 2021).
Personalization: Integration of imaging, tumor biology (Ki67, TILs, subtype), and morphology substantially increases accuracy and utility of treatment stratification (Malhaire et al., 21 Nov 2025).

The credible technical progress in radiomics, deep learning, transformer architectures, topology-informed modeling, and causal inference substantiates pCR as a powerful clinical and research endpoint—central to the evolution of response-adaptive, precision oncology in breast cancer.