Prediction of pathological complete response to neoadjuvant therapy in breast cancer integrating intratumoral and peritumoral delta radiomics with clinical features: a study based on multiparametric MRI

Abstract

Background

To develop a combined model integrating intratumoral and peritumoral delta-radiomics from multi-parametric MRI with clinical features for predicting pathological complete response (pCR, i.e., Miller-Payne grade V response) to neoadjuvant chemotherapy (NAT) in breast cancer.

Methods

A total of 254 patients with breast cancer from two hospitals were retrospectively included. Radiomics features were extracted from both intratumoral and peritumoral regions on multi-parametric MRI obtained at baseline and after the second cycle of NAT, and delta radiomics features were subsequently derived to quantify longitudinal changes between the two time points. Six machine learning algorithms were employed to construct and compare delta-radiomics and clinical models to identify the optimal algorithm. The integrated model combining intratumoral and peritumoral delta radiomics and clinical features was developed based on the optimal algorithm to predict pCR. Additionally, the Shapley Additive Explanations (SHAP) method was applied to interpret the contributions of features within the optimal model.

Results

Compared with other machine learning algorithms, ExtraTrees algorithm achieved superior predictive performance for both delta radiomics and clinical models. Receiver operating characteristic and decision curve analysis demonstrated that the multimodal model outperformed the delta radiomics and clinical models in predicting pCR. Furthermore, SHAP analysis revealed that dynamic changes in intratumoral and peritumoral radiomics features during neoadjuvant therapy jointly drive the predictive performance of the fusion model.

Conclusions

This study highlighted the potential of intratumoral and peritumoral regions in delta radiomics analysis to evaluate neoadjuvant therapy for breast cancer, thereby further broadening the prospects for personalized decision-making in breast cancer treatment.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02545-3.

Introduction

Breast cancer is the malignant tumor with the highest incidence and mortality rates among all female cancers [1]. Neoadjuvant therapy (NAT) is a systemic treatment administered prior to radical surgery for primary breast cancer, which has demonstrated high efficacy in tumor size reduction and downstaging, and is established as the standard treatment for locally advanced breast cancer [2–4]. The optimal therapeutic endpoint of NAT is defined as pathologic complete response (pCR), that the attainment of pCR is significantly associated with improved long-term prognosis [5, 6]. However, previous studies revealed that pCR rates following NAT were only 19% across patient populations [7]. For patients with poor response to NAT, they not only endure the adverse effects of treatment but also face delayed surgical opportunities, thereby increasing the risk of tumor progression [8, 9]. Thus, the early prediction of pCR probability is critical for optimizing individualized treatment strategies.

In clinical practice, dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) has been established as the most accurate imaging modality for assessing NAT response in breast cancer patients [10–12]. Furthermore, numerous studies have integrated medical imaging with machine learning by employing high-throughput feature extraction and analytical techniques from medical images, which demonstrated the capability to identify imperceptible patterns beyond human visual perception and play an important role in predicting prognostic outcomes for breast cancer patients [10, 11].

As an emerging radiomics approach, delta radiomics focuses on the dynamic tracking of changes in lesion characteristics over time, which has demonstrated transformative potential in oncological practice [12]. Some studies have validated the significant clinical utility of delta radiomics in predicting therapeutic response to NAT [13, 14]. Current pCR prediction models predominantly focus on intratumoral heterogeneity, yet the peritumoral normal tissue is susceptible to microscopic invasion or dissemination by malignant cells, resulting in subtle but detectable structural alterations in breast cancer [15, 16]. Therefore, radiomics features extracted from the peritumoral region might reflect critical information of tumor heterogeneity and prognosis [17]. However, little is known about the potential value of delta radiomics features derived from intratumoral and peritumoral regions in predicting response to NAT in patients with breast cancer.

This study aims to develop and validate the application value of multiparametric MRI delta radiomics features derived from intratumoral and different peritumoral regions for predicting response to NAT. Furthermore, combined models including delta radiomics features and clinical markers are evaluated to determine the predictive performance for pCR (i.e., Miller-Payne grade V response) following NAT to develop a more promising tool for personalized therapeutic decision-making in breast cancer management.

Materials and methods

The retrospective study protocol received ethical approval from institutional review boards of two participating institutions: Institution A (QYFY WZLL 29981) and Institution B (No.2023 − 352), which waived the requirement for written informed consent. This study complies with the Declaration of Helsinki and was performed according to ethics committee approval.

Patients

We retrospectively collected data from 254 breast cancer patients across two independent institutions between October 2018 and October 2023. The inclusion criteria were defined as follows: (1) histopathological confirmation of breast cancer through needle biopsy; (2) baseline DCE-MRI and DCE-MRI following two cycles of NAT were performed; (3) all participants underwent surgical procedures after NAT, with concomitant pathological evaluation of therapeutic response.

The exclusion criteria were established as follows: (1) patients with therapeutic interventions including radiotherapy, chemotherapy, or surgical procedures before baseline examination; (2) occult breast carcinoma or bilateral breast malignancies; (3) lesions demonstrating complete radiological response following two cycles of NAT, confirming tumor eradication in postoperative histopathology.

The study population was stratified as follows: Institution A contributed 183 cases, which were randomly allocated into training (n = 128) and internal validation (n = 55) cohorts at a 7:3 ratio, and Institution B provided an external validation cohort comprising 71 cases. Figure 1 depicted the inclusion and exclusion criteria for patients and cohort distribution.

Fig. 1.

Patient enrollment flowchart

Data analysis

Clinicopathological data analysis

The clinical and pathological information was collected from medical records of patients (Fig. 2a), including age, menopausal status, lymph node metastasis status, estrogen receptor (ER) status, expression levels of progesterone receptor (PR) status, human epidermal growth factor receptor-2 (HER-2) status, Ki-67 proliferation index and molecular subtypes. The definitions of PR, ER, HER2 positivity and molecular subtypes were based on the recommendations in the College of American Pathologists Guidelines [18]. The treatment efficacy of NAT was assessed using the Miller-Payne (MP) histological grading system. Patients with MP V grade were divided into pCR group, while patients with MP I-IV grades were divided into nonpCR group.

Fig. 2.

Workflow of the radiomics analysis

Conventional imaging analysis

All multiparametric MRI images from Institutions A and B were independently retrieved by two experienced breast radiologists at the Picture Archiving and Communication System workstation. Quantitative imaging analysis was also performed. Any discrepancies in image interpretation were resolved through consensus discussion. The evaluated imaging features included tumor location, axial T2-weighted imaging (T2WI) signal intensity, time signal intensity curve (TIC) pattern, background parenchymal enhancement (BPE) level, minimum apparent diffusion coefficient (ADC) mapping value, average ADC mapping value, gland density on MRI images before NAT.

Additionally, volumes of the tumors were assessed both before (pre-NAT) and after two cycles of NAT (post-2nd NAT). The volumes of tumors were calculated on first-phase contrast-enhanced scans by 3D Slicer Segment Statistics (version 4.11; https://www.slicer.org/). Volume% = (Volume_{pre NAT} − Volume_{post−2nd NAT}) / Volume_{pre NAT}×100.

Imaging scans

Both Institutions A and B performed MRI examinations using 3.0T scanners (GE Discovery MR 750w and Siemens Magnetom Prisma) equipped with 8-channel phased-array breast coils. Patients were positioned prone with both breasts naturally suspended within the coil aperture. Our study used imaging protocols including T2WI, diffusion-weighted imaging (DWI), ADC mapping and DCE sequences. For DCE-MRI, gadobutrol (Gadovist; Bayer Schering Pharma) was administered as the contrast agent at a standard dose of 0.1mmol/kg body weight (0.1 mL/kg) through a peripheral intravenous catheter placed in the dorsal hand vein. The contrast agent was injected at 2 mL/s using a power injector, followed by a 20 mL saline flush at the same injection rate. Detailed imaging parameters for both scanners are summarized in Table 1.

Table 1.

The routine MRI sequences and scanning parameters of the two institutions

Institution	Sequence	FOV (mm)	Matrix	TR (ms)	TE (ms)	Flip angle (°)	Section Thickness (mm)	intersection gap (mm)	b-value (s/mm2)
A	DCE	360 × 360	360 × 360	4.3	2.2	15	1	0	-
	T2WI	320 × 320	320 × 320	7260	85	111	4	1	-
	DWI	320 × 320	128 × 128	6378	69.8	-	4	1	1000
B	DCE	360 × 360	320 × 288	6.6	1.7	8	2	0	-
	T2WI	360 × 360	320 × 320	5400	85	111	5	1	-
	DWI	320 × 320	96 × 128	6000	70	-	6	0	1000

FOV: field of view; TR: time of repetition; TE: time of echo

Image segmentation and preprocessing

Multiparametric MRI images acquired from pre-NAT and post-2nd NAT were imported into 3D Slicer software (version 4.11; https://www.slicer.org/) for tumor segmentation. Tumor segmentation was performed by a breast radiologist with 5 years of experience, who manually delineated region of interest (ROI)s slice-by-slice on T2WI, first-phase contrast-enhanced scans, DWI, and ADC mapping. The radiologist was blinded to the pathologic findings.

Image preprocessing included image resampling to achieve isotropic voxels (1 × 1 × 1mm³) to standard spatial resolution and intensity normalization to minimize inter-scanner variability. The original tumor mask was defined as the intratumoral ROI. To characterize the peritumoral microenvironment, the intratumoral mask was radially dilated outward by 3mm, 5mm, and 7mm, respectively. The corresponding peritumoral ROIs were generated by subtracting the original intratumoral mask from each dilated mask, thereby yielding three annular peritumoral regions surrounding the tumor. Figure 2b demonstrated an example of the ROI segmentation and mask dilation process. Image preprocessing was implemented on OnekeyAI platform (https://github.com/OnekeyAI-Platform/onekey).

To evaluate interobserver variability, another radiologist with 8 years of breast imaging experience independently re-segmented ROIs in a randomly selected subset of 30 patients from three cohorts one month after initial segmentation. Interobserver agreement was quantified using the intraclass correlation coefficient (ICC), with an ICC > 0.75 indicating good consistency.

Radiomics feature extraction and selection

The radiomics features were extracted using the radiomics software package Pyradiomics version 3.10 (detailed code is provided in the supplementary materials). The PyRadiomics extraction parameters were predefined as follows: bin width = 25, resampled pixel spacing = (1.0,1.0,1.0), interpolator = sitkBSpline, normalization enabled, normalize scale = 100, remove outliers = 3, correct mask = True, force2D = False, distances = [1], voxel array shift = 300, and weighting norm = manhattan. Wavelet features were generated using the coif1 wavelet decomposition, and LoG features were extracted with sigma values of 1.0, 2.0, 3.0, 4.0, and 5.0mm. When necessary, ROI masks were resampled to match image geometry using nearest-neighbor interpolation. Notably, radiomics features were extracted separately from the intratumoral ROI and from each peritumoral ROI. A total of 1,316 radiomics features were extracted from each ROI on both pre-NAT and post-2nd NAT images, respectively. The extracted features were categorized into nine classes: (1) 3D shape features, n = 14; (2) original first-order features, n = 18; (3) original gray-level co-occurrence matrix features, n = 24; (4) original gray-level run length matrix features, n = 16; (5) original gray-level size zone matrix features, n = 16; (6) original gray-level dependence matrix features, n = 14; (7) original neighboring gray-tone difference matrix features, n = 5; (8) wavelet-transformed features, n = 744; and (9) Laplacian of Gaussian transformed features, n = 465. The wavelet-transformed features were derived from eight wavelet decompositions, with 93 radiomics features extracted from each decomposition. The LoG-transformed features were extracted using five sigma values, namely 1.0, 2.0, 3.0, 4.0, and 5.0mm, with 93 features extracted for each sigma level.

The delta radiomics features are derived by calculating the differences between the post-2nd NAT features values and the pre-NAT features values:

All delta radiomics features were processed using the Z-score standardization method. The process of radiomics features selection involved the following four steps: (1) radiomics features with p < 0.05 were reserved in the next analysis by using the Mann-Whitney U test; (2) the Pearson’s correlation analysis was performed between pairwise radiomics features within the training set. Features demonstrating intercorrelation coefficients exceeding 0.8 underwent exclusion through randomised selection to mitigate multicollinearity; (3) the maximum relevance–minimum redundancy (mRMR) method was applied to further select radiomics features, and the 10 features with the greatest relevance and least redundancy were retained. (4) least absolute shrinkage selection operator (LASSO) methodology was implemented for radiomics feature selection, employing regularized regression to identify statistically significant predictors while mitigating overfitting. The final feature subset selected from the training set was fixed and directly applied to the validation set without additional feature selection.

Model construction

We performed model construction in three stages. In the first phase, delta_tumor radiomics and clinical models were developed using six machine learning algorithms (including Extremely Randomized Trees (ExtraTrees), Random Forest (RF), Support Vector Machine (SVM), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), Multilayer Perceptron (MLP)) across T2WI, first-phase DCE-MRI, DWI, and ADC sequences. These models were assessed at internal validation cohort to determine the most effective algorithm. In the second phase, multimodal models were developed using the optimal algorithm identified in the first phase. Four multimodal models were constructed by integrating independent clinical predictors with multi-parametric delta radiomics features derived from the intratumoral region and different peritumoral regions, including delta_tumor+clinical, delta_tumor+3mm+clinical, delta_tumor+5mm+clinical, delta_tumor+7mm+clinical models. These models were compared in the training and internal validation cohorts to identify the optimal multimodal model. The model with the favorable predictive performance in the internal validation cohort was selected for further evaluation. Finally, the selected optimal multimodal model was compared with the corresponding delta radiomics model and clinical model across the training, internal validation, and external validation cohorts using ROC analysis, DCA, and evaluation metrics (i.e., accuracy, sensitivity, specificity, precision, brier score, and DCA Mean Net Benefit). The workflow of our study was shown in Fig. 2.

Statistical analysis

Clinical and pathological features were screened using a rigorous, training-only feature-selection pipeline to prevent data leakage and ensure the independence of the internal validation and external validation cohorts. First, baseline characteristics were compared across the training, internal validation and external validation cohorts to assess distributional consistency; however, no variables were excluded based on this comparison. Feature selection was performed strictly within the training cohort. Specifically, variables were compared between the outcome groups (pCR vs. non-pCR) within the training set. Continuous variables were analyzed using the independent-samples t-test or one-way analysis of variance (ANOVA) when normally distributed with homogeneous variances; otherwise, the Mann–Whitney U test or Kruskal–Wallis test was applied. Categorical variables were compared using the chi-square test or Fisher’s exact test. Variables showing significant differences between outcome groups in the training set (p < 0.05) were selected as candidate predictors. These candidates were then subjected to univariable logistic regression analysis. Features with p < 0.05 in the univariable analysis were subsequently entered into a multivariable logistic regression model to identify independent predictors and construct the clinical model.

The diagnostic performance of the models was assessed using the area under the receiver operating characteristic curve (AUC), together with sensitivity, specificity, and accuracy. Statistical comparisons of ROC curves were performed using DeLong’s test. Decision curve analysis (DCA), Brier score, and mean net benefit derived from DCA were employed to further evaluate the clinical utility of the models. A two-tailed p-value < 0.05 was considered statistically significant for all analyses, except in the multivariable logistic regression, where a stricter threshold of p < 0.01 was applied to identify independent predictors. All statistical analyses were conducted using R software (detailed code is provided in the supplementary materials).

Model explanation and feature visualization

SHapley Additive exPlanations (SHAP) is a sophisticated interpretability framework rooted in cooperative game theory, specifically designed to allocate the contribution of each feature in a predictive model with precision [19]. By utilizing Shapley values, SHAP rigorously assesses the marginal contribution of each feature across all possible permutations of input combinations. The method was deployed to show the feature importance of machine learning models in our study (supplementary materials for detailed codes).

Results

Repeatability analysis

Intraclass correlation coefficients were employed to evaluate the inter-observer consistency of the feature extraction [20]. The results demonstrated that the ICC values for all delta radiomics features extracted from the 30 patients ranged from 0.752 to 1.000 (all > 0.75), indicating favorable reproducibility consistency.

Study population

A total of 254 patients were included in our study (Institution A, 183 patients; Institution B, 71 patients). The mean age of patients with pCR (i.e., Miller-Payne grade V response) (Institution A, 51 patients; Institution B, 25 patients) was 50 (13.5) years. In the patients with non-pCR (Institution A, 132 patients; Institution B, 46 patients), the mean age was 51.5 (13.75) years. The clinical and imaging features of all patients from two institutions were demonstrated in Table 2.

Table 2.

Clinical and radiological characteristics in the three cohorts

Characteristics	Training (n = 128)			Internal validation (n = 55)			External validation (n = 71)			P_Group
	Non-pCR (n = 94)	pCR (n = 34)	P	Non-pCR (n = 38)	pCR (n = 17)	P	Non-pCR (n = 46)	pCR (n = 25)	P	P
Age	49.27 ± 9.33	49.29 ± 9.32	0.988	49.03 ± 8.26	51.71 ± 10.52	0.361	53.39 ± 9.75	46.24 ± 9.32	0.004*	0.522
Average ADC mapping value	0.91 (0.84, 1.04)	0.96 (0.86, 1.13)	0.153	0.92 (0.83, 1.06)	1.02 (0.89, 1.20)	0.160	0.87 (0.77, 0.93)	0.92 (0.81, 1.04)	0.084	0.015*
Minimum ADC mapping value	0.66 (0.56, 0.78)	0.71 (0.62, 0.88)	0.020*	0.73 (0.59, 0.82)	0.71 (0.67, 0.76)	0.560	0.73 (0.67, 0.78)	0.78 (0.69, 0.88)	0.100	0.002*
Volume%	62.73 (47.71, 79.77)	84.91 (74.73, 94.20)	< 0.001*	74.02 (60.02, 90.67)	88.94 (83.62, 90.56)	0.043*	62.85 (52.71, 80.94)	91.73 (82.18, 96.81)	< 0.001*	0.065
Volumepre NAT	48.69 (21.38, 103.71)	44.44 (15.54, 80.95)	0.277	54.38 (19.08, 148.52)	43.20 (24.57, 77.66)	0.834	24.70 (10.12, 36.64)	17.55 (9.36, 34.94)	0.563	< 0.001*
Volumepost−2nd NAT	16.28 (5.29, 31.96)	3.95 (0.72, 10.51)	< 0.001*	8.78 (4.02, 30.23)	5.62 (3.47, 14.79)	0.228	6.30 (2.75, 11.91)	1.72 (0.84, 3.84)	< 0.001*	< 0.001*
Lymph node metastasis status			0.388			0.742			0.223	0.424
Negative	24 (25.53%)	12 (35.29%)		9 (23.68%)	5 (29.41%)		7 (15.22%)	7 (28.00%)
Positive	70 (74.47%)	22 (64.71%)		29 (76.32%)	12 (70.59%)		39 (84.78%)	18 (72.00%)
Menstrual status			0.638			0.679			0.026*	0.210
Menopausal	45 (47.87%)	14 (41.18%)		15 (39.47%)	5 (29.41%)		19 (41.30%)	18 (72.00%)
Not menopausal	49 (52.13%)	20 (58.82%)		23 (60.53%)	12 (70.59%)		27 (58.70%)	7 (28.00%)
TIC pattern			1.000			1.000			0.377	< 0.001*
Plateau	10 (10.64%)	4 (11.76%)		6 (15.79%)	3 (17.65%)		23 (50.00%)	16 (64.00%)
Washout	84 (89.36%)	30 (88.24%)		32 (84.21%)	14 (82.35%)		23 (50.00%)	9 (36.00%)
ER status			0.008*			0.411			0.041*	0.097
Negative	22 (23.40%)	17 (50.00%)		14 (36.84%)	9 (52.94%)		7 (15.22%)	10 (40.00%)
Positive	72 (76.60%)	17 (50.00%)		24 (63.16%)	8 (47.06%)		39 (84.78%)	15 (60.00%)
PR status			< 0.001*			0.524			0.226	0.598
Negative	24 (25.53%)	22 (64.71%)		15 (39.47%)	9 (52.94%)		14 (30.43%)	12 (48.00%)
Positive	70 (74.47%)	12 (35.29%)		23 (60.53%)	8 (47.06%)		32 (69.57%)	13 (52.00%)
HER-2 status			0.039*			0.006*			< 0.001*	< 0.001*
Negative	40 (42.55%)	7 (20.59%)		21 (55.26%)	2 (11.76%)		39 (84.78%)	11 (44.00%)
Positive	54 (57.45%)	27 (79.41%)		17 (44.74%)	15 (88.24%)		7 (15.22%)	14 (56.00%)
Ki-67			0.799			0.416			0.120	0.672
Negative	13 (13.83%)	6 (17.65%)		6 (15.79%)	1 (5.88%)		11 (23.91%)	2 (8.00%)
Positive	81 (86.17%)	28 (82.35%)		32 (84.21%)	16 (94.12%)		35 (76.09%)	23 (92.00%)
Gland density			0.965			0.381			0.461	< 0.001*
Entirely fatty	1 (1.06%)	0 (0.00%)		0 (0.00%)	0 (0.00%)		4 (8.70%)	2 (8.00%)
Scattered fibroglandular	11 (11.70%)	3 (8.82%)		4 (10.53%)	0 (0.00%)		21 (45.65%)	10 (40.00%)
Heterogeneously dense	67 (71.28%)	26 (76.47%)		29 (76.32%)	16 (94.12%)		20 (43.48%)	10 (40.00%)
Extremely dense	15 (15.96%)	5 (14.71%)		5 (13.16%)	1 (5.88%)		1 (2.17%)	3 (12.00%)
BPE level			0.697			0.545			1.000	< 0.001*
Minimal	32 (34.04%)	15 (44.12%)		16 (42.11%)	5 (29.41%)		37 (80.43%)	21 (84.00%)
Mild	33 (35.11%)	12 (35.29%)		9 (23.68%)	7 (41.18%)		7 (15.22%)	3 (12.00%)
Moderate	15 (15.96%)	4 (11.76%)		8 (21.05%)	4 (23.53%)		2 (4.35%)	1 (4.00%)
Marked	14 (14.89%)	3 (8.82%)		5 (13.16%)	1 (5.88%)		0 (0.00%)	0 (0.00%)
T2WI signal intensity			0.219			0.519			0.189	< 0.001*
Hypointensity	46 (48.94%)	22 (64.71%)		19 (50.00%)	11 (64.71%)		1 (2.17%)	3 (12.00%)
Isointensity	30 (31.91%)	9 (26.47%)		9 (23.68%)	4 (23.53%)		8 (17.39%)	6 (24.00%)
Hyperintensity	18 (19.15%)	3 (8.82%)		10 (26.32%)	2 (11.76%)		37 (80.43%)	16 (64.00%)
Molecular subtypes			0.050			0.322			0.190	< 0.001*
Luminal A	11 (11.70%)	3 (8.82%)		5 (13.16%)	1 (5.88%)		8 (17.39%)	1 (4.00%)
Luminal B	64 (68.09%)	16 (47.06%)		20 (52.63%)	10 (58.82%)		32 (69.57%)	17 (68.00%)
HER2 over expression	17 (18.09%)	13 (38.24%)		8 (21.05%)	6 (35.29%)		1 (2.17%)	2 (8.00%)
Triple negative	2 (2.13%)	2 (5.88%)		5 (13.16%)	0 (0.00%)		5 (10.87%)	5 (20.00%)
Tumor location			0.613			0.329			0.153	0.613
Outer upper quadrant	51 (54.26%)	14 (41.18%)		21 (55.26%)	10 (58.82%)		25 (54.35%)	7 (28.00%)
Outer lower quadrant	8 (8.51%)	4 (11.76%)		3 (7.89%)	2 (11.76%)		5 (10.87%)	6 (24.00%)
Inner upper quadrant	6 (6.38%)	4 (11.76%)		3 (7.89%)	1 (5.88%)		5 (10.87%)	5 (20.00%)
Inner lower quadrant	12 (12.77%)	4 (11.76%)		2 (5.26%)	3 (17.65%)		4 (8.70%)	1 (4.00%)
Area centralis	17 (18.09%)	8 (23.53%)		9 (23.68%)	1 (5.88%)		7 (15.22%)	6 (24.00%)

Volume%: relative change in volume during neoadjuvant therapy; TIC: time signal intensity curve; ER: estrogen receptor; PR: progesterone receptor; HER-2: human epidermal growth factor receptor-2; BPE: background parenchymal enhancement

Clinicopathological and radiographic characteristics

Univariate analysis revealed that ER status, PR status, minimum ADC mapping value, Volume_{pre NAT}, HER-2 and Volume% were significantly associated with the response to NAT (p < 0.05). No significant differences were observed for the remaining clinicopathological and imaging features between the pCR and non-pCR groups (p > 0.05). Subsequent multivariate logistic regression analysis identified minimum ADC mapping value and Volume% as independent predictors of pCR following NAT (Table 3).

Table 3.

Univariate and multivariate logistic regression analysis of clinical and radiological characteristics

Characteristic	Univariate analysis			Multivariate analysis
	OR	95%CI	P-value	OR	95%CI	P-value
ER status	3.273	1.438–7.546	0.005*	0.559	0.148–1.951	0.373
PR status	5.347	2.344–12.759	< 0.001*	5.244	1.525–19.833	0.010
Volume%	0.949	0.922–0.972	< 0.001*	0.954	0.922–0.982	0.003*
Volumepost−2nd NAT	1.024	1.005–1.052	0.046*	1.003	0.988–1.028	0.789
HER-2 status	0.350	0.130–0.846	0.026*	0.543	0.172–1.604	0.278
Minimum ADC mapping value	0.614	0.414–0.888	0.012*	0.500	0.295–0.803	0.006*

ER: estrogen receptor; PR: progesterone receptor; Volume%: relative change in volume during neoadjuvant therapy; OR: odds ratio; HER-2: human epidermal growth factor receptor-2

Diagnostic performance of delta radiomics and clinical models based on six machine learning algorithms in evaluating pCR

Delta radiomics and clinical models were developed using six machine learning algorithms. In the internal validation cohort, the ExtraTrees algorithm demonstrated favorable and relatively stable predictive performance compared to the other five classifiers. Specifically, the radiomics-based ExtraTrees model achieved the highest AUC (0.803) and the highest decision curve analysis (DCA) mean net benefit (0.120) in internal validation. Furthermore, the clinical ExtraTrees model yielded the highest accuracy (0.782) and the lowest Brier score (0.191) among all clinical models. (Fig. 3; Table 4). DeLong test revealed that in the internal validation set, the ExtraTrees clinical model showed statistically significant differences compared to the LightGBM (p = 0.008) clinical model. DCA substantiated that the ExtraTrees models offered promising clinical utility (Fig. 4). Consequently, ExtraTrees was identified as the optimal algorithm for this study and was subsequently employed for the construction of multi-modality predictive models.

Fig. 3.

ROC curves of delta_tumor radiomics and clinical models built by six machine learning algorithms in training and internal validation validation cohorts. (a-b) ROC curves for delta_tumor radiomics models in training and internal validation cohorts; (c-d) ROC curves for clinical models in training and internal validation cohorts

Table 4.

Comparison of AUCs in the different models

	AUC(95%CI)
	ExtraTrees	RF	SVM		XGBoost		LightGBM		MLP
clinical
training	0.849(0.773–0.925)	0.956(0.923–0.989)	0.794(0.708–0.881)		0.983(0.965-1.000)		0.885(0.827–0.943)		0.764(0.668–0.860)
internal validation	0.719(0.575–0.864)	0.657(0.492–0.823)	0.749(0.614–0.885)		0.659(0.503–0.814)		0.618(0.451–0.784)		0.656(0.510–0.802)
delta tumor
training	0.969(0.937-1.000)	0.985(0.965-1.000)	0.849(0.849–0.925)		0.997(0.990-1.000)		0.974(0.950–0.999)		0.918(0.859–0.977)
internal validation	0.803(0.679–0.928)	0.796(0.660–0.933)	0.720(0.557–0.863)		0.755(0.615–0.894)		0.738(0.590–0.885)		0.665(0.504–0.827)

AUC: area under the curve; ExtraTrees: Extremely Randomized Trees; RF: Random Forest; SVM: Support Vector Machine; XGBoost: eXtreme Gradient Boosting; LightGBM: Light Gradient Boosting Machine; MLP: Multilayer Perceptron

Fig. 4.

Decision curve analysis of delta_tumor radiomics and clinical models built by six machine learning algorithms in training and internal validation cohorts. (a-b) Decision curve analysis for delta_tumor radiomics models in training and internal validation cohorts; (c-d) Decision curve analysis for clinical models in training and internal validation cohorts

Diagnostic performance of multimodal models in evaluating pCR

Figure 5 presents the ROC curves and DCA for the multimodal models integrating intratumoral and peritumoral (3, 5, and 7mm) features in training and internal validation cohorts. The ROC analysis demonstrated that the delta_tumor+5mm+clinical model achieved superior predictive performance in the internal validation cohort (AUC = 0.847). In the internal validation set, DCA showed that the delta_tumor+5mm+clinical model provided relatively favorable net benefit within the predefined clinically relevant threshold range. Given that the delta_tumor+5mm+clinical model demonstrated the best discriminative performance among the four fusion models in the internal validation cohort, the delta_tumor+5mm+clinical model was selected as the optimal multimodal model for further comparison with the corresponding delta radiomics model and the clinical model.

Fig. 5.

ROC and decision curve analysis of four multimodal models in the training and internal validation cohorts. (a-b) ROC curves for four multimodal models in training, internal validation, and external validation cohorts; (c-d) Decision curve analysis for four multimodal models in training, internal validation, and external validation cohorts

Thus, a comparative analysis was performed among the delta_tumor+5mm radiomics model, clinical model and delta_tumor+5mm+clinical model across training, internal validation and external validation cohorts. The results indicated that the AUC values of delta_tumor+5mm+clinical model were higher than other models across training, internal validation and external validation cohorts (Fig. 6). DCA showed that the delta_tumor+5mm+clinical model provided relatively favorable net benefit within the predefined clinically relevant threshold range compared with delta_tumor+5mm radiomics and clinical models (Fig. 6). Other evaluation metrics (i.e., accuracy, sensitivity, specificity, precision, brier score, and DCA Mean Net Benefit) for other models are shown in Tables 5 and 6.

Fig. 6.

Comparison of delta radiomics, clinical and multimodal models in training, internal validation, and external validation cohorts. (a-c) ROC curves for delta radiomics, clinical and fusion models in training, internal validation, and external validation cohorts; (d-f) Decision curve analysis for delta radiomics, clinical and fusion models in training, internal validation, and external validation cohorts

Table 5.

Performance of the multimodal model across training and internal validation cohorts using the ExtraTrees algorithm

	Accuracy	Sensitivity	Specificity		Precision	AUC (95%CI)	Brier score	DCA Mean Net Benefit
delta tumor +clinical
training	0.898	0.971	0.872	0.733		0.968 (0.938–0.998)	0.090	0.109
internal validation	0.800	0.824	0.789	0.636		0.831 (0.724–0.938)	0.166	0.127
delta tumor+3mm +clinical
training	0.977	0.971	0.979	0.943		0.996 (0.990-1.000)	0.049	0.119
internal validation	0.764	0.824	0.737	0.583		0.808 (0.694–0.922)	0.171	0.123
delta tumor+5mm +clinical
training	0.934	0.941	0.936	0.842		0.973 (0.941-1.000)	0.083	0.112
internal validation	0.800	0.765	0.816	0.650		0.847 (0.745–0.948)	0.155	0.124
delta tumor+7mm +clinical
training	0.906	0.882	0.915	0.789		0.970 (0.945–0.995)	0.073	0.107
internal validation	0.800	1.000	0.711	0.607		0.808 (0.694–0.922)	0.178	0.124

ExtraTrees: Extremely Randomized Trees

Table 6.

Performance of delta radiomics, clinical, and multimodal models across training, internal validation, and external validation cohorts using the ExtraTrees algorithm

	Accuracy	Sensitivity	Specificity	Precision		AUC (95%CI)		Brier score	DCA Mean Net Benefit
clinical
training	0.828	0.706	0.872	0.667		0.849 (0.773–0.925)	0.148		0.093
internal validation	0.782	0.529	0.895	0.692		0.719 (0.575–0.864)	0.191		0.114
external validation	0.775	0.800	0.761	0.645		0.793 (0.685–0.901)	0.186		0.157
deltatumor+5mm radiomics
training	0.953	0.853	0.989	0.967		0.965 (0.928-1.000)		0.082	0.109
internal validation	0.764	0.882	0.711	0.577		0.830 (0.718–0.941)		0.157	0.122
external validation	0.761	0.640	0.826	0.667		0.760 (0.648–0.872)		0.249	0.146
delta tumor+5mm +clinical
training	0.934	0.941	0.936	0.842		0.973 (0.941-1.000)		0.083	0.112
internal validation	0.800	0.765	0.816	0.650		0.847 (0.745–0.948)		0.155	0.124
external validation	0.746	0.760	0.739	0.613		0.799 (0.695–0.903)		0.234	0.151

ExtraTrees: Extremely Randomized Trees

Model interpretation

SHAP analysis was performed to interpret the delta_tumor+5mm+clinical model. The SHAP beeswarm and bar plots showed that model predictions were mainly driven by delta-radiomics features, with Delta_wavelet-LHL_glszm_SmallAreaLowGrayLevelEmphasis_C, Delta_log-sigma-3-0-mm-3D_glcm_MaximumProbability_C, and Delta_wavelet-HLL_glcm_Idmn_T2 showing the greatest contribution to model output (Fig. 7a, b). Volume% and Minimum_ADC_value pre-NAT contributed to a lesser extent. The representative waterfall plots further demonstrated how these features jointly influenced individual predictions towards pCR and non-pCR outcomes (Fig. 7c, d). Overall, these results suggest that dynamic, multi-region radiomics changes during NAT, rather than a single baseline imaging biomarker, underpinned the predictive performance of the fusion model.

Fig. 7.

Feature contributions to pCR prediction in the delta_tumor+5mm+clinical model based on SHAP analysis

Discussion

In this study, multi-parametric MRI delta radiomics models were developed and compared using six machine learning algorithms. Subsequently, a multimodal model integrating intratumoral and peritumoral delta radiomic features with clinicopathologic and radiological indicators was established for the early prediction of pCR in breast cancer patients following two cycles of NAT. Our results demonstrate that the multimodal model incorporating both intratumoral and peritumoral information exhibits high predictive performance and clinical utility, highlighting the importance of the peritumoral microenvironment in assessing NAT efficacy. Furthermore, the comprehensive multimodal model demonstrated higher and more stable predictive efficacy for pCR than the model based on the ExtraTrees algorithm alone. These findings suggest that the model integrating intratumoral and peritumoral delta radiomics with clinical features provides a reliable tool for the early evaluation of NAT response and holds great potential in supporting individualized therapeutic decision-making for breast cancer patients.

NAT has been established as the standard treatment for locally advanced breast cancer [7]. Recent imaging studies have focused on prediction of the response to NAT using delta radiomics model in patients with breast cancer. For instance, Guo et al. developed the delta radiomics model utilizing DCE-MRI data acquired before and after the first cycle of neoadjuvant chemotherapy (NAC), demonstrating excellent predictive performance for pCR with AUC of 0.917 and 0.842 in the training and validation cohorts [13]. Furthermore, Fan et al. investigated delta radiomics features derived from DCE-MRI images during the second NAC cycle and achieved good predictive result (AUC = 0.726) [14]. These collective findings suggested that delta radiomics based on DCE-MRI was a promising tool for prediction of pCR. In our study, we found that delta radiomics machine learning models based on multiparametric MRI had also high predictive performance for response to NAT in the training and internal validation cohorts (e.g., AUC = 0.997, 0.803). Nevertheless, we also noted that delta radiomics models built solely on intratumoral ROIs showed limited and inconsistent performance across all six machine learning algorithms in the internal validation cohort (AUCs of 0.666–0.803). Thus, solely extracting delta radiomics features from intratumoral region might be insufficient to reflect the dynamic changes in the tumor parenchyma and stroma associated with tumor heterogeneity and aggressiveness.

The peritumoral region of breast cancer exhibits significant microstructural alterations during tumor cell invasion and dissemination [21]. Braman et al. demonstrated that machine learning models combined pre-NAT intratumoral and peritumoral radiomics features had higher performance than single-region models (i.e., intratumoral or peritumoral region) in predicting pCR in HER2-positive breast cancer patients [17]. Moreover, Niu et al. suggested that molecular subtype assessment models incorporating both intratumoral and peritumoral radiomics features outperformed single-region models in diagnostic efficacy [22]. Our study implemented a systematic approach to achieve radiomics analysis of peritumoral regions. Automated volumetric expansions of ROIs were performed at incremental distances of 3mm, 5mm, and 7mm from the tumor margin to construct delta_tumor+3mm+clinical model, delta_tumor+5mm+clinical model and delta_tumor+7mm+clinical model. The comparison of multimodal models across different spatial extents revealed that models integrating both intratumoral and peritumoral regions had superior diagnostic performance, with the delta_tumor+5mm+clinical model demonstrating the highest predictive stability. The finding suggests that the addition of peritumoral features improves the predictive performance of our models. Additionally, a limited peritumoral margin may be sufficient to capture pCR-related microenvironmental signals while reducing irrelevant background information from more distant tissues [23, 24]. Thus, delta radiomics features of peritumoral region might be potential markers for pathological alterations surrounding the breast tumor mass, such as peritumor lymphovascular invasion, peritumoral lymphocytic infiltration and peritumoral edema [25, 26].

Radiomics can analyze comprehensive features of tissue heterogeneity through multidimensional data integration to construct predictive models for clinical outcomes [27]. However, the methodology has persistent validation challenges regarding feature reproducibility and model interpretability. SHAP provides radiomics features with clinical interpretability through feature contribution decomposition algorithms, and its clinical application has been confirmed in many previous studies [28, 29]. This study presented a SHAP-based visual analytics framework for reverse decomposition of predictive models, which significantly enhanced the interpretability of black-box machine learning architectures through quantitative evaluation of radiomic feature weight distribution patterns [30]. In our study, the delta_tumor+5mm+clinical model was predominantly driven by delta radiomics features rather than by a single baseline imaging biomarker. Specifically, Delta_wavelet-LHL_glszm_SmallAreaLowGrayLevelEmphasis_C, Delta_log-sigma-3-0-mm-3D_glcm_MaximumProbability_C, and Delta_wavelet-HLL_glcm_Idmn_T2 contributed most to the model output, whereas Volume% and Minimum_ADC_value pre-NAT contributed to a lesser extent. Biologically, these texture-based features collectively quantify intratumoral heterogeneity by characterizing the spatial distribution and local uniformity of voxel intensities within the tumor. Their changes during NAT therefore capture the progressive loss of tumor heterogeneity and the emergence of more homogeneous tissue that accompany favorable therapeutic response, a pattern repeatedly linked to treatment sensitivity in prior radiomics studies. Notably, the peritumoral feature (i.e., Delta_wavelet-LHL_glszm_ZoneEntropy_DWI_5mm) also contributed to pCR prediction, indicating that information from the surrounding microenvironment may complement intratumoral changes in the fusion model. Overall, dynamic changes in intratumoral and peritumoral radiomics features during neoadjuvant therapy jointly drive the predictive performance of the fusion model, supporting intratumoral and peritumoral delta radiomics as a robust and biologically grounded strategy for non-invasive pCR prediction in breast cancer. It is also noted that our model excludes patients achieving early radiologic complete response due to segmentation technicalities, potentially omitting the most treatment-sensitive subgroup. In contrast, baseline minimum ADC was the only pre-NAT feature among significant features. Higher baseline minimum ADC values were generally associated with positive SHAP values, suggesting a tendency toward greater predicted tumor shrinkage. This confirms that the pretreatment state fundamentally dictates the therapeutic trajectory, consistent with the high chemotherapy sensitivity observed in highly proliferative, diffusion-restricted tumors [31].

Previous studies have indicated that some imaging and clinical markers including ER, PR, HER-2, Ki-67 index, and Volume% could be used to predict pCR [32, 33]. Our results demonstrated that Volume% and the pre-NAT minimum ADC value were independent predictors of pCR following NAT, which was consistent with previous research findings. Qian et al. constructed a nomogram model based on age, the American Joint Committee on Cancer T stage, Ki-67 index, HER2, and HR status, and it confirmed its non-invasive applicability for personalized prediction of pCR in breast cancer patients who underwent NAT prior to surgery through both internal and external validation [34]. However, our results of clinical models were unstable. For instance, the clinical models based on LightGBM and MLP achieved AUC values of 0.618 and 0.656 in the internal validation set, which might suggest that clinical information is insufficient in capturing the heterogeneity of tumor pathophysiological alterations or may be attributable to the relatively small sample size. The high-dimensional imaging features might contribute to acquiring more detailed information about the tumor that cannot be detected easily by the naked eye and compensate for the limitations of small sample sizes, hence multimodal models might contribute to the comprehensive visualization of tumor-associated transformations across multiple biological scales, from macroscopic architectural changes to microscopic molecular alterations [35]. Liu et al. conducted a multicenter retrospective study to investigate the potential of combining mpMRI radiomic features with clinical information for predicting the likelihood of pCR in breast cancer. The results indicated that the predictive performance of the combined model (AUC = 0.86) was superior to that of the radiomic model alone (AUC = 0.79) [11]. Herrero et al. showed that the combination of MRI perfusion parameters, radiomic features, and clinical variables significantly enhances the accuracy of predicting pCR [36]. Our study further corroborated that multimodal models achieved significantly higher predictive performance than clinical or delta radiomics models in the machine learning framework.

In addition, to ensure a rigorous interpretation of our research findings, it is essential to fully consider the scientific context and boundary conditions of this study. While our model demonstrates high internal validity, its generalizability to broader clinical scenarios involves several nuanced considerations. First, this model was developed exclusively for patients with measurable residual disease after two cycles of NAT. Five patients with early complete radiological response were excluded at enrolment. This was a deliberate clinical design choice: early complete responders have an excellent prognosis and have no need for predictive testing, while patients with residual disease represent the high-uncertainty population where this tool adds actual clinical value. This model should not be extrapolated to unselected patients starting NAT. Second, pCR was defined in this study as MP V grade. This differs from the common alternative definition of ypT0/is ypN0, and direct quantitative comparison to studies using this endpoint should be made with appropriate caution.

Our study demonstrated promising predictive performance in assessment of the response to NAT in patients with breast cancer. Nevertheless, some limitations should be acknowledged. First, our study had relatively limited sample size and the imbalance between pCR and non-pCR cases. A larger, multi-institutional study will be implemented to enhance the generalizability of our findings in future. Second, our study had various molecular subtypes of breast cancer. Therefore, subsequent research should build predictive models to investigate treatment responses of subtype-specific breast cancer. Additionally, the pCR definition used here differs from the standard ypT0/is ypN0, potentially limiting direct comparisons with other large-scale trials. Future studies should employ standardized pathological criteria to ensure consistency and facilitate cross-study validation. Moreover, while manual ROI segmentation demonstrated high interobserver consistency (ICC > 0.75), the approach may introduce operator-dependent variability. Implementation of semi-automated segmentation algorithms could improve reproducibility and reduce subjective bias in future studies.

Conclusions

In this study, we developed a promising model for pCR in patients with residual measurable disease after two cycles of neoadjuvant chemotherapy. Our results demonstrated that delta radiomics features, derived from longitudinal changes in intratumoral and peritumoral characteristics, carry valuable information regarding tumor biology and should be incorporated in future radiomics studies. By integrating delta radiomics with clinically independent predictive factors, the multi-modality model achieved improved predictive performance and generalizability. Moreover, SHAP-based feature importance analysis enhanced model interpretability, facilitating its potential translation into clinical practice and ultimately supporting timely treatment adjustment and personalized therapeutic strategies.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

ADC

Apparent diffusion coefficient

AUC

Area under the curve

BPE

Background parenchymal enhancement

DCA

Decision Curve Analysis

DCE-MRI

Dynamic contrast enhanced magnetic resonance imaging

DWI

Diffusion-weighted imaging

ER

Estrogen receptor

ExtraTrees

Extremely Randomized Trees

HER-2

Human epidermal growth factor receptor-2

ICC

Intraclass correlation coefficient

LightGBM

Light Gradient Boosting Machine

LASSO

Least absolute shrinkage selection operator

MLP

Multilayer Perceptron

M-P

Miller-Payne

mRMR

Maximum relevance–minimum redundancy

NAC

Neoadjuvant chemotherapy

NAT

Neoadjuvant therapy

pCR

Pathologic complete response

PR

Progesterone receptor

RF

Random Forest

ROC

Receiver operating characteristic curve

ROI

Region of interest

SHAP

SHapley Additive exPlanations

SVM

Support Vector Machine

TIC

Time signal intensity curve

T2WI

Axial T2-weighted imaging

XGBoost

eXtreme Gradient Boosting

Author contributions

Conceptualization: LX, ZHC and CJJ. Data curation: FM and HYX Formal analysis: HJ and LJJ. Funding acquisition: CJJ. Investigation: MXN and MXR Methodology: ZHC. Project administration: CJJ. Resources: LX and CJJ. Software: ZHC. Supervision: CJJ. Validation: FM and MXR. Visualization: ZHC and HH. Writing-original draft: LX, ZHC, HH, LJJ, HYX, MXN, FM and HYX Writing-review & editing: LX, ZHC MXR and CJJ.

Funding

This work was supported by the National Natural Science Foundation of China [grant numbers 82072004].

Data availability

The datasets generated and/or analysed during the current study are not publicly available due [REASON WHY DATA ARE NOT PUBLIC] but are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the Declaration of Helsinki and was approved by the Institutional Review Boards of both participating institutions (Institution A: Medical Ethics Committee of the Affiliated Hospital of Qingdao University, approval No. QYFY WZLL 29981; Institution B: Medical Ethics Committee of the Affiliated Yantai Yuhuangding Hospital of Qingdao University, approval No. 2023 − 352). The requirement for written informed consent was waived by the Institutional Review Boards due to the retrospective nature of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.