Binary and Ternary Classification Prediction for Breast Cancer and Breast Sclerosing Adenosis With Interpretable Artificial Intelligence From Clinical and Imaging Features: A Retrospective, Diagnostic Accuracy Cohort Study

ABSTRACT

Background

Sclerosing adenosis (SA) and breast cancer (BC) often exhibit overlapping clinical, imaging, and pathological characteristics, making them difficult to differentiate. SA may also coexist with BC (SA + BC), including ductal carcinoma in situ (SA‐DCIS) and invasive breast cancer (SA‐IBC), which complicates diagnosis even when core‐needle biopsy (CNB) suggests SA. This study aimed to develop interpretable AI‐based binary and ternary classification models that leverage clinical and imaging features to distinguish SA‐only from SA + BC and to further differentiate among SA‐only, SA‐DCIS, and SA‐IBC.

Methods

We retrospectively analyzed a cohort of 726 patients with SA (January 2006 to December 2021), comprising 537 SA‐only and 189 SA + BC cases (90 SA‐DCIS, 99 SA‐IBC). Multiple machine learning algorithms—logistic regression, support vector machine, decision tree, XGBoost, and random forest—were compared using AUC, accuracy, F1‐score, and C‐index. Model interpretability was assessed with SHAP to elucidate feature contributions and identify key predictors. Additionally, we incorporated an independent external validation cohort consisting of 113 patients to verify the model's effectiveness.

Results

XGBoost consistently outperformed other algorithms in both tasks. Eight features emerged as most informative: age, ultrasound BI‐RADS category, maximum and minimum ultrasound diameters, ultrasound margin characteristics, biopsy procedure, mammographic density, and microcalcifications. For binary classification (SA‐only vs. SA + BC), XGBoost achieved an AUC of 0.925, accuracy of 0.883, and C‐index of 0.844. For ternary classification (SA‐only, SA‐DCIS, SA‐IBC), the model achieved an AUC of 0.888, accuracy of 0.811, and C‐index of 0.813. Age, ultrasound BI‐RADS, and minimum lesion diameter were consistently top predictors. We further proposed a three‐tier interpretability framework (global, cohort‐level; local, subgroup‐level; and individual, case‐level) to facilitate clinical translation.

Conclusion

Given the substantial risk of coexisting of SA with DCIS or IBC, and the potential for CNB to underestimate disease due to limited sampling, lesions diagnosed as SA on CNB should be evaluated with additional modalities before determining the need for surgical excision. The proposed interpretable AI model enhances discrimination between SA‐only and SA with concomitant breast cancer (SA + BC), thereby supporting more informed clinical decision‐making in breast disease management.

XGBoost model (AUC 0.925/0.888) using 8 clinical‐ultrasound features accurately classified SA‐only versus SA + BC (including DCIS/IBC), outperforming other algorithms; proposed interpretability framework enhances clinical decision‐making for SA management.

Abbreviations

ADH

atypical ductal hyperplasia

AI

artificial intelligence

AUC

area under the ROC curve

BC

breast cancer

CNB

core‐needle biopsy

DCA

decision curve analysis

DCIS

ductal carcinoma in situ

DT

decision tree

FFB

fast frozen biopsy

IBC

invasive breast cancer

IHC

immunohistochemistry

MG

mammography

ML

machine learning

MRI

magnetic resonance imaging

RF

random forest

ROC curve

receiver operating characteristic curve

SA

sclerosing adenosis

SHAP

Shapley additive explanations

US

ultrasound

XGBoost

eXtreme Gradient boosting

1. Introduction

According to recent authoritative statistics, breast cancer (BC) has become the most frequently diagnosed malignancy and the leading cause of cancer‐related death among women worldwide [1, 2, 3]. Sclerosing adenosis (SA) is a common benign histopathologic finding characterized by lobulocentric proliferation of epithelial and myoepithelial cells, disordered acinar architecture, and stromal fibrosis. More than 25% of biopsies for benign breast disease contain SA, and women with SA have approximately a twofold increased risk of subsequent BC [4, 5, 6]. A transcriptomic study of SA‐containing biopsies identified an SA‐specific genetic signature associated with future BC development [7], and a biomarker panel has been proposed to predict BC risk in women with SA [8].

Differentiating SA from BC is challenging based on clinical presentation, physical examination, imaging, and even intraoperative frozen section [9, 10]. Further complicating diagnosis, SA can mimic BC as a palpable mass or a solid lesion on imaging, and it may also coexist with BC, including ductal carcinoma in situ (DCIS) and invasive BC (IBC) [11, 12, 13]. Ultrasound (US), mammography (MG), and magnetic resonance imaging (MRI) have shown suboptimal sensitivity and specificity for distinguishing SA from BC and for differentiating SA‐only from SA + BC [12, 14, 15, 16, 17, 18]. At present, no imaging modality or predictive model reliably differentiates SA‐only from SA + BC, particularly when stratifying among SA‐only, SA + DCIS, and SA + IBC.

With the rapid advances in machine learning (ML) and the availability of large‐scale clinical datasets, artificial intelligence (AI) has been increasingly applied to differential diagnosis from medical imaging, individualized outcome prediction, and treatment decision support [19, 20, 21]. Radiomics, which extracts quantitative features from medical images, further enables disease diagnosis, risk stratification, and clinical assessment [22, 23]. For example, Kha et al. developed an efficient MRI radiomics signature to predict 1p/19q co‐deletion status in lower‐grade gliomas [24], and Le et al. proposed a sequence‐based ML model with two‐level feature selection to predict protein crystallization propensity [25]. In BC, deep learning–based radiomics has shown promise for early detection, histologic grade estimation, axillary lymph node status prediction in early‐stage disease, and survival assessment [26, 27].

Given the high dimensionality and large feature spaces inherent to radiomics and multimodal datasets, feature selection methods are essential to identify the most informative variables and reduce dimensionality while preserving predictive signal [24, 25]. Moreover, interpretable ML approaches—such as Shapley additive explanations (SHAP)—are increasingly employed to elucidate model outputs and provide clinically understandable, case‐level and cohort‐level explanations, thereby facilitating transparent and trustworthy deployment in practice [24].

Regarding the differential diagnosis of SA and BC, several studies have explored deep learning and radiomics approaches. Liang et al. developed and validated a nomogram incorporating patient age and US breast imaging reporting and data system (BI‐RADS) to distinguish malignant tumors from SA among lesions with the same BI‐RADS category, thereby aiding selection of the most suspicious mass for biopsy [13]. Another study reported that combining US BI‐RADS mapping with breast elastography can effectively distinguish SA from BC [28]. However, these efforts primarily address differentiation between SA and BC at the lesion level, whereas our work targets the clinically consequential task of identifying coexistent malignancy in patients with SA—specifically differentiating SA‐only from SA + BC and further stratifying SA + DCIS versus SA + IBC.

In this study, we employed multiple machine learning algorithms to develop binary (SA‐only vs. SA + BC) and ternary (SA‐only vs. SA + DCIS vs. SA + IBC) predictive models to detect malignancy among patients with SA.

2. Patients and Methods

2.1. Ethics Approval

Ethical approval was granted by the Ethics Committee of Peking Union Medical College Hospital, Chinese Academy of Medical Sciences(approval number: K3610). Written informed consent was waived due to the anonymized nature of the data.

2.2. Patient Cohort

We retrospectively identified 767 consecutive patients with SA who underwent surgery at PUMC Hospital between January 2006 and December 2021, including 567 SA‐only and 200 SA + BC cases. Coexistence (SA + BC) required histologic spatial contiguity of the two lesions. Patients with missing or incomplete clinicopathologic parameters were excluded. Overall, 41 of 767 patients (5.6%) were excluded: 36/573 (6.3%) from the SA‐only group and 5/194 (2.6%) from the SA + BC group. Given the low exclusion rate, imputation was not performed. The final analysis cohort comprised 726 patients: 537 SA‐only and 189 SA + BC (90 SA + DCIS and 99 SA + IBC) (Figure 1).

Figure 1.

The processing flow of study.

Inclusion criteria were: (1) female sex; (2) age 18–89 years; and (3) unilateral SA confirmed by histopathology. Among eligible patients, those who were pregnant or had contralateral breast cancer were excluded.

To enhance temporal generalizability and provide external validation, we assembled an independent validation cohort of 113 newly diagnosed patients from January 2022 to September 2025, applying the same inclusion and exclusion criteria. This cohort included 73 SA‐only and 40 SA + BC cases, of which 12 were SA + DCIS and 28 were SA + IBC.

Clinical characteristics, imaging features from US and MG reports, and pathologic findings were abstracted by investigators. Immunohistochemistry (IHC) for ER and PR was reported as the percentage of tumor cell nuclei staining positive. Prior to 2010, ER/PR positivity was defined as ≥ 10% nuclear staining; from 2010 onward, ≥ 1% was considered positive [29]. For human epidermal growth factor receptor 2 (HER2), IHC 2+ cases underwent fluorescence in situ hybridization (FISH). Before 2013, a HER2/CEP17 ratio ≥ 2.2 was positive, 1.8‐2.2 equivocal, and < 1.8 negative; from 2014 onward, a ratio ≥ 2.0 was positive and < 2.0 negative [30].

2.3. Nomogram Models

We constructed two nomograms to predict binary (SA‐only vs. SA + BC) and ternary (SA‐only vs. SA + DCIS vs. SA + IBC) outcomes using the same analytic pipeline. Candidate predictors from clinical and imaging variables were first screened by univariate logistic regression, and significant factors were entered into multivariable logistic regression. Forest plots were used to display odds ratios (ORs) and p‐values for significant predictors.

Nomograms were developed with the rms package in R to provide a visual tool for individualized risk estimation. Discrimination was assessed by the area under the receiver operating characteristic curve (AUC), with ROC curves used for visualization. Calibration accuracy was evaluated using calibration curves generated by the calibrate function, and clinical utility was examined via decision curve analysis (DCA) using the decision_curve function.

To evaluate temporal generalizability, we performed external validation in an independent cohort of 113 cases collected from another hospital between January 2022 and August 2025, applying the same inclusion criteria and modeling procedures.

2.4. Feature Selection and Unbalanced Processing

We implemented an integrated feature‐selection strategy that combined ML screening with clinical nomogram evidence. First, six ML methods—recursive feature elimination (RFE) [31], decision tree (DT) [32], random forest (RF) [33], chi‐square test [25], Lasso [34] and eXtreme Gradient boosting (XGBoost) [35]—were used for preliminary selection. A voting scheme aggregated results across methods by tallying the selection frequency of each feature to establish a priority ranking; features chosen by at least three methods were designated high‐priority candidates. Next, we incorporated features that appeared in both the binary and ternary clinical nomograms to ensure alignment with variables highlighted by clinical modeling. The final feature set emphasized consensus variables between the ML voting results and the nomograms, while also adding key predictors present in both nomograms but not captured by the ML screen. The target size of the final set was approximately 50% of the original feature pool (eight features), determined by empirical performance and clinical expert input. This multi‐algorithm, clinically informed approach was designed to enhance robustness and interpretability. Given class imbalance and the risk of bias toward majority classes, we applied the Synthetic Minority Over‐sampling Technique (SMOTE) for data augmentation. Oversampling parameters were tuned empirically to balance minority‐class representation while limiting noise introduction, based on cross‐validated predictive performance.

2.5. Model Establishment and Evaluation Index

The model classifier needs to be chosen according to the research question, size of dataset, type of data and the evaluation index. In this study, we chose logistic regression (LR) [36], DT [37], RF [38], support vector machine (SVM) [39], linear support vector machine (LSVM) [40], and XGBoost [41] as model classifiers. The predictive performances of the models were evaluated with precision, sensitivity, F1‐score, area under the ROC curve (AUC), accuracy and C‐index [42, 43]. The calculation of these metrics was based on the classification of the confusion matrix [44].

The hyperparameter optimization process was designed to minimize the logarithmic loss (log loss), which provides a stringent evaluation of the quality of predicted probabilities. To ensure robustness and mitigate overfitting, model performance was evaluated using stratified fivefold cross‐validation on the training set (constituting 80% of the total data). An exhaustive grid search was employed across predefined hyperparameter ranges. These search ranges were determined based on established practices in the literature and preliminary exploratory experiments. The detailed search spaces and the final optimized parameters for all models are summarized in Supporting Information: Table 6.

The predictive performance of the above six model classifiers was evaluated and compared for binary (SA‐only vs. SA + BC) and ternary (SA‐only vs. SA + DCIS vs. SA + IBC) predictions.

2.6. Shapley Additive Explanation (SHAP) Interpretability Analysis

Model interpretations were generated using SHAP by transforming model features into clinical variables, providing explanations for the machine learning classification model by calculating the contribution of each feature to the prediction results [45, 46]. The global and local contributions of eight feature variables respectively in binary and ternary classification prediction were performed according to the SHAP values. Individualized SHAP maps for patients with SA, SA + BC, atypical ductal hyperplasia (ADH), SA + DCIS and SA + IBC were presented to show how each feature impact the prediction result.

2.7. Statistical Analysis

Python (3.10) and R (4.0.3) software were utilized for statistical analysis of the data. Continuous variables were described by mean plus or minus standard deviation (SD), while categorical variables by frequency and percentage (%). Continuous variables were tested by t‐test whereas categorical variables by Fisher's or chi‐square test for the comparisons. Statistical significance level was set at two‐sided with p‐value < 0.05.

3. Results

3.1. Patient Characteristics

Among the 537 SA‐only patients, 55 (10.2%) had precancerous lesions—41 with ADH and 14 with lobular carcinoma in situ (LCIS). The most common benign coexisting lesions were fibroadenoma (FA, 33.1%), apocrine metaplasia (AM, 29.4%), and intraductal papilloma (IDP, 26.8%). Among SA + BC patients who underwent intraoperative fast frozen biopsy (FFB), diagnostic uncertainty occurred in 10 of 53 SA + DCIS cases (18.9%) and 14 of 71 SA + IBC cases (19.7%) (Supporting Information: Table 2).

Compared with SA‐only patients (n = 537; 74.0%), those with SA + BC (n = 189; 26.0%) were significantly older and more likely to exhibit the following features: MG micro‐calcifications; MG BI‐RADS categories 4B‐5; US BI‐RADS categories 4B‐5; ill‐defined US margins; irregular US morphology; multifocality; larger lesions with greater maximum and minimum US diameters; hyperechogenicity on US; disordered, increased US vascularity; and abnormal axillary lymph nodes on US (Table 1). No significant differences were observed in MG‐detected nodules/masses, US diameter ratio (min/max), or US aspect ratio (Table 1).

Table 1.

The statistical description of the variables of SA‐only and SA + BC.

Clinicopathological and imaging characteristics	SA‐only N = 537 (%)	SA + BC N = 189 (%)	p‐value
Age	42 ± 12 a	47 ± 11 a	< 0.001
Age group
< 40	205 (38.2)	28 (14.8)	< 0.001
40–49	238 (44.3)	85 (44.4)
50–59	68 (12.7)	58 (30.7)
≥ 60	26 (4.8)	19 (10.1)
Mammography (MG) features
MG lesion BI‐RADS			< 0.001
0–4a	488 (90.9)	130 (68.8)
4b–5	49 (9.1)	59 (31.2)
MG density			0.007
Fatty	30 (5.6)	21 (11.1)
Scattered	160 (29.8)	64 (33.9)
Heterogeneous	300 (55.9)	97 (51.3)
Dense	47 (8.8)	7 (3.7)
MG micro‐calcification			< 0.001
No	435 (81.0)	144 (60.3)
Yes	102 (19.0)	75 (39.7)
MG nodule/mass			0.456
No	201 (37.4)	65 (34.4)
Yes	336 (62.6)	124 (65.6)
Ultrasound (US) features
US lesion BI‐RADS			< 0.001
0–4a	397 (73.9)	36 (19.0)
4b–5	140 (26.1)	153 (81.0)
US aspect ratio			0.968
Normal	509 (94.8)	179 (74.7)
Abnormal	28 (5.2)	10 (5.3)
US boundary			< 0.001
Clear	333 (62.0)	51 (27.0)
Unclear	204 (38.0)	138 (73.0)
US diameter (Median ± IQR)
max (cm)	1.4 ± 0.6 a	1.7 ± 1.3 a	< 0.001
min (cm)	0.6 ± 0.3 a	1.0 ± 0.7 a	< 0.001
ratio (min/max)	0.56 ± 0.31 a	0.60 ± 0.32 a	0.053
US morphology			< 0.001
Regular	262 (48.8)	43 (22.8)
Irregular	275 (51.2)	146 (77.2)
US hyperechoicity			< 0.001
No	437 (81.4)	110 (58.2)
Yes	100 (18.6)	79 (41.8)
US blood flow			< 0.001
Normal	241 (44.9)	52 (27.5)
Peripheral	100 (18.6)	35 (18.5)
Internal	152 (28.3)	55 (29.1)
Rich/disordered	44 (8.2)	47 (24.9)
US lymph node			< 0.001
Normal	520 (96.8)	154 (81.5)
Mild abnormal	13 (2.4)	20 (10.6)
Abnormal	4 (0.7)	15 (7.9)

Abbreviations: BC, breast cancer; BI‐RADS, breast imaging reporting and data system; IQR, interquartile range; MG, mammography; SA, sclerosing adenosis.

^a This data was reported as median ± IQR.

Across SA (n = 537) versus SA + DCIS (n = 90) versus SA + IBC (n = 99), there were stepwise increases in the proportions of the following: MG BI‐RADS 4B‐5, MG microcalcifications, US BI‐RADS 4B‐5, ill‐defined US margins, maximum and minimum US diameters, irregular US morphology, US hyperechogenicity, disordered US blood flow, and abnormal axillary lymph nodes on US (Table 2).

Table 2.

The statistical description of the variables of SA‐only, SA + DCIS and SA + IBC.

Clinicopathological and imaging characteristics	SA‐only N = 537 (%)	SA + DCIS N = 90 (%)	SA + IBC N = 99 (%)	p‐value
Age	42 ± 12 a	48 ± 13 b	48 ± 10 a	< 0.001
Age group				< 0.001
< 40	205 (38.2)	17 (18.9)	11 (11.1)
40–49	238 (44.3)	35 (38.9)	49 (49.5)
50–59	68 (12.7)	29 (32.2)	29 (29.3)
≥ 60	26 (4.8)	9 (10.0)	10 (10.1)
Mammography (MG) features
MG lesion BI‐RADS				< 0.001
0–4a	488 (90.9)	70 (77.8)	60 (60.6)
4b–5	49 (9.1)	20 (22.2)	39 (39.4)
MG density				0.009
Fatty	30 (5.6)	13 (14.4)	8 (8.1)
Scattered	160 (29.8)	26 (28.9)	38 (38.4)
Heterogeneous	300 (55.9)	48 (53.3)	49 (49.5)
Dense	47 (8.8)	3 (33.0)	4 (4.0)
MG micro‐calcification				< 0.001
No	435 (81.0)	59 (65.6)	55 (55.6)
Yes	102 (19.0)	31 (34.4)	44 (44.4)
MG nodule/mass				0.727
No	201 (37.4)	30 (33.3)	35 (35.4)
Yes	336 (62.6)	60 (66.7)	64 (64.6)
Ultrasound(US) features
US lesion BI‐RADS				< 0.001
0–4a	397 (73.9)	22 (24.4)	14 (14.1)
4b–5	140 (26.1)	68 (75.6)	85 (85.9)
US aspect ratio				0.531
Normal	509 (94.8)	87 (96.7)	92 (92.9)
Abnormal	28 (5.2)	3 (3.3)	7 (7.1)
US boundary				< 0.001
Clear	333 (62.0)	33 (36.7)	18 (18.2)
Unclear	204 (38.0)	57 (63.3)	81 (81.8)
US diameter
max (cm)	1.40 ± 0.60 a	1.70 ± 1.33 a	1.80 ± 1.40 a	< 0.001
min (cm)	0.60 ± 0.30 a	0.90 ± 0.69 a	1.10 ± 0.90 a	< 0.001
ratio (min/max)	0.56 ± 0.31 a	0.53 ± 0.32 a	0.63 ± 0.18 b	0.012
US morphology				< 0.001
Regular	262 (48.8)	27 (30.0)	16 (16.2)
Irregular	275 (51.2)	63 (70.0)	83 (83.8)
US hyperechoicity				< 0.001
No	437 (81.4)	61 (67.8)	49 (49.5)
Yes	100 (18.6)	29 (32.2)	50 (50.5)
US blood flow				< 0.001
Normal	241 (44.9)	33 (36.7)	19 (19.2)
Peripheral	100 (18.6)	14 (15.6)	21 (21.1)
Internal blood	152 (28.3)	22 (24.4)	33 (33.3)
Rich/disordered	44 (8.2)	21 (23.3)	26 (26.3)
US lymph node				< 0.001
Normal	520 (96.8)	84 (93.3)	70 (70.7)
Mild abnormal	13 (2.4)	4 (4.4)	16 (16.2)
Abnormal	4 (0.7)	2 (2.2)	13 (13.1)

Abbreviations: BI‐RADS, breast imaging reporting and data system; DCIS, ductal carcinoma in situ; IBC, invasive breast cancer; IQR, interquartile range; MG, mammography; SA, sclerosing adenosis; SD, standard deviation.

^a This data was reported as Median ± IQR.

^b This data was reported as mean ± SD.

In the external validation cohort of 113 patients (2022–2025), those with SA + BC (n = 40) were significantly older and more likely to present with higher BI‐RADS categories (4B‐5), microcalcifications, nodules/masses, ill‐defined ultrasound margins, irregular morphology, hyperechogenicity, abnormal lymph nodes, and larger lesion diameters compared with SA‐only patients (n = 73) (all p < 0.05) (Table 3). Further stratification of SA + BC into SA + DCIS (n = 12) and SA + IBC (n = 28) showed a stepwise increase—from SA‐only to SA + DCIS to SA + IBC—in age, high BI‐RADS categories, microcalcifications, nodules/masses, and malignant ultrasound features (ill‐defined margins, irregular morphology, hyperechogenicity, and nodal abnormalities) (all p < 0.05). These findings align imaging phenotypes with pathological severity and corroborate trends observed in the primary cohort (Table 4).

Table 3.

The statistical description of the variables of SA‐only and SA + BC (temporal validation cohort).

Clinicopathological and imaging characteristics	SA‐only N = 73 (64.6%)	SA + BC N = 40 (35.4%)	p‐value
Age	42.63 ± 8.78 a	50.00 ± 8.98 a	< 0.001
Age group
< 40	26 (35.6)	4 (10.0)	< 0.001
40–49	31 (42.5)	12 (30.0)
50–59	11 (15.1)	18 (45.0)
≥ 60	5 (6.8)	6 (15.0)
Mammography (MG) features
MG lesion BI‐RADS			< 0.001
0–4a	61 (83.6)	10 (25.0)
4b–5	12 (16.4)	30 (75.0)
MG density			0.965
Fatty	0 (0.0)	0 (0.0)
Scattered	8 (10.9)	5 (12.5)
Heterogeneous	58 (79.5)	31 (77.5)
Dense	7 (9.6)	4 (10.0)
MG micro‐calcification			< 0.001
No	66 (90.4)	17 (42.5)
Yes	7 (9.6)	23 (57.5)
MG nodule/mass			< 0.01
No	31 (42.5)	6 (15.0)
Yes	42 (57.5)	34 (85.0)
Ultrasound (US) features
US lesion BI‐RADS			< 0.001
0–4a	61 (83.6)	10 (25.0)
4b–5	12 (16.4)	30 (75.0)
US aspect ratio			0.242
Normal	70 (95.9)	36 (90.0)
Abnormal	3 (4.1)	4 (10.0)
US boundary			< 0.001
Clear	31 (42.5)	4 (10.0)
Unclear	42 (57.5)	36 (90.0)
US diameter
max (cm)	1.57 ± (0.76) a	2.60 ± (1.39) a	< 0.001
min (cm)^	0.70 ± (0.50) b	1.20 ± (0.70) b	< 0.001
ratio (min/max)	0.60 ± (0.26) b	0.56 ± (0.20) a	0.572
US morphology			< 0.05
Regular	16 (21.9)	1 (2.5)
Irregular	57 (78.1)	39 (97.5)
US hyperechoicity			< 0.001
No	46 (63.0)	9 (22.5)
Yes	27 (37.0)	31 (77.5)
US blood flow			0.599
Normal	24 (32.9)	9 (22.5)
Peripheral	12 (16.4)	6 (15.0)
Internal	20 (27.4)	12 (30.0)
Rich/disordered	17 (23.3)	13 (32.5)
US lymph node			< 0.01
Normal	67 (91.8)	26 (65.0)
Mild abnormal	4 (5.5)	8 (20.0)
Abnormal	2 (2.7)	6 (15.0)

Abbreviations: BC, breast cancer; BI‐RADS, breast imaging reporting and data system; DCIS, ductal carcinoma in situ; IQR, interquartile range; MG, mammography; SA, sclerosing adenosis; SD, standard deviation

^a This data was reported as Mean ± SD.

^b This data was reported as Median ± IQR.

Table 4.

The statistical description of the variables of SA‐only, SA + DCIS and SA + IBC (temporal validation cohort).

Clinicopathological and imaging characteristics	SA‐only N = 73 (64.6%)	SA + DCIS N = 12 (10.6%)	SA + IBC N = 28 (24.8%)	p‐value
Age	42.00 ± (12.00) a	47.25 ± (12.31) b	51.18 ± (7.07) b	< 0.001
Age group				< 0.01
< 40	26 (35.6)	3 (25.0)	1 (3.6)
40–49	31 (42.5)	2 (16.7)	10 (35.7)
50–59	11 (15.1)	5 (41.7)	13 (46.4)
≥ 60	5 (6.8)	2 (16.7)	4 (14.3)
Mammography (MG) features
MG lesion BI‐RADS				< 0.001
0–4a	69 (94.5)	3 (25.0)	7 (25.0)
4b–5	4 (5.5)	9 (75.0)	21 (75.0)
MG density				0.516
Fatty	0 (0.0)	0 (0.0)	0 (0.0)
Scattered	8 (10.9)	0 (0.0)	5 (17.9)
Heterogeneous	58 (79.5)	10 (83.3)	21 (75.0)
Dense	7 (9.6)	2 (16.7)	2 (7.1)
MG micro‐calcification				< 0.001
No	66 (90.4)	5 (41.7)	12 (42.9)
Yes	7 (9.6)	7 (58.3)	16 (57.1)
MG nodule/mass				< 0.05
No	31 (42.5)	2 (16.7)	4 (14.3)
Yes	42 (57.5)	10 (83.3)	24 (85.7)
Ultrasound(US) features
US lesion BI‐RADS				< 0.001
0–4a	61 (83.6)	5 (41.7)	5 (17.9)
4b–5	12 (16.4)	7 (58.3)	23 (82.1)
US aspect ratio				0.106
Normal	70 (95.9)	12 (100)	24 (85.7)
Abnormal	3 (4.1)	0 (0.0)	4 (14.3)
US boundary				< 0.01
Clear	31 (42.5)	1 (8.3)	3 (10.7)
Unclear	42 (57.5)	11 (91.7)	25 (89.3)
US diameter
max (cm)	1.40 ± (1.10) a	2.61 ± (1.52) b	2.60 ± (1.37) b	< 0.001
min (cm)	0.70 ± (0.50) a	1.07 ± (0.49) b	1.37 ± (0.61) b	< 0.001
ratio (min/max)	0.56 ± (0.25) a	0.51 ± (0.24) b	0.58 ± (0.18) b	0.202
US morphology				< 0.05
Regular	16 (48.8)	0 (0.0)	1 (3.6)
Irregular	57 (51.2)	12 (100)	27 (96.4)
US hyperechoicity				< 0.001
No	46 (63.0)	5 (41.7)	4 (14.3)
Yes	27 (37.0)	7 (58.3)	24 (85.7)
US blood flow				0.883
Normal	24 (32.9)	3 (25.0)	6 (21.4)
Peripheral	12 (16.4)	2 (16.7)	4 (14.3)
Internal blood	20 (27.4)	4 (33.3)	8 (28.6)
Rich/disordered	17 (23.3)	3 (25.0)	10 (35.7)
US lymph node				< 0.001
Normal	67 (91.8)	8 (66.7)	18 (64.3)
Mild abnormal	4 (5.5)	4 (33.3)	4 (14.3)
Abnormal	2 (2.7)	0 (0.0)	6 (21.4)

Abbreviations: BI‐RADS, breast imaging reporting and data system; DCIS, ductal carcinoma in situ; IBC, invasive breast cancer; IQR, interquartile range; MG, mammography; SA, sclerosing adenosis; SD, standard deviation.

^a This data was reported as Median±IQR.

^b This data was reported as Mean ± SD.

Beyond these core similarities, the SA + BC patients in the external cohort presented with a more pronounced “classic” malignant imaging profile. Specifically, they showed a higher proportion of MG BI‐RADS 4b‐5 lesions, a greater prevalence of microcalcifications and masses/nodules on mammography, and significantly larger lesion sizes on ultrasound. These differences likely stem from a combination of factors inherent to real‐world validation, such as temporal effects, inter‐institutional variations in clinical practice, and potential differences in patient referral patterns or population structure.

3.2. Establishment and Validation of Nomogram Models to Predict Binary and Ternary Categorized SA

For the binary nomogram (SA‐only vs. SA + BC), selected predictors included age, US BI‐RADS, maximum US diameter, US diameter ratio (min/max), US aspect ratio, US margin (boundary), biopsy/surgery procedure, MG density, and MG micro‐calcifications (Supporting Information: Figure 1A). The 726 patients were split 4:1 into training (n = 581) and testing (n = 145) sets (Supporting Information: Table 3). In the training set, the model achieved an accuracy of 0.861, AUC of 0.913, and C‐index of 0.913; corresponding metrics in the testing set were 0.834, 0.868, and 0.868, respectively (Supporting Information: Figures 1B,C). Calibration curves showed good agreement between predicted and observed outcomes in both sets (Supporting Information: Figure 1D,E). DCA indicated greater net benefit than “treat‐all” or “treat‐none” strategies in both training and testing cohorts (Supporting Information: Figure 1F,G).

For the ternary nomogram (SA‐only vs. SA + DCIS vs. SA + IBC), predictors included age, US BI‐RADS, maximum and minimum US diameters, US margin, biopsy/surgery procedure, MG BI‐RADS, MG density, and MG micro‐calcifications (Supporting Information: Figure 2A). Performance in the training set was accuracy 0.800, macro‐averaged AUC 0.819, and C‐index 0.772; in the testing set, accuracy 0.793, macro‐averaged AUC 0.783, and C‐index 0.730 (Supporting Information: Figure 2B,C). Calibration plots demonstrated favorable concordance between predicted and observed class probabilities in both sets (Supporting Information: Figure 2D,E).

3.3. Integrated Features Selection by AI and Nomogram and Processing of Imbalanced Data

The top five clinical and imaging features identified by each method are summarized in Supporting Information: Table 4. Of the 16 candidate variables, eight were selected by at least one method and were prioritized by selection frequency (Supporting Information: Table 5). Among the seven predictors shared by both the binary and ternary nomograms, six—age, US BI‐RADS, maximum US diameter, US margin (boundary), biopsy procedure, and MG micro‐calcifications—also ranked within the top seven features in the ML‐based screening. To preserve clinically important variables captured by the nomograms, we additionally included MG density, which appeared in both nomograms but was not consistently selected by ML. This yielded a final set of eight key features (50% of the original 16) for AI model development (Supporting Information: Table 5).

Sampling parameters for class imbalance were tuned to balance effectiveness against overfitting. Overly aggressive oversampling risks noise inflation, whereas conservative settings may insufficiently correct imbalance. Based on empirical validation, the oversampling ratios were set to 1.5 for the binary task and 4:1:1 for the ternary task.

3.4. Binary and Ternary Classification AI Prediction

We evaluated six classifiers—LR, SVM, LSVM, DT, RF, and XGBoost—using the eight selected features for both binary and ternary tasks. XGBoost delivered the best overall performance, achieving the highest AUC (binary: 0.925; ternary: 0.888), accuracy (binary: 0.883; ternary: 0.811), and C‐index (binary: 0.844; ternary: 0.813) (Figure 2a,b; Tables 5 and 6). Calibration curves indicated closer agreement between predicted probabilities and observed outcomes for XGBoost than for other benchmarks (Figure 3a). DCA demonstrated greater net benefit for XGBoost across clinically relevant threshold probabilities compared with “treat‐all” and “treat‐none” strategies in the test set (Figure 2c).

Figure 2.

(a) ROC curves for triple classification of each model. (b) ROC curve for triple classification of XGBoost. (c) DCA of the XGBoost model in the test cohort to predict SA and SA + BC(N = 145). (d) DCA of the XGBoost model in the temporal validation cohort to predict SA and SA + BC(N = 113). AUC, area under the ROC curve; DCA, decision curve analysis; DT, decision tree; LR, logistic regression; LSVM, linear support vector machine; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, eXtreme Gradient boosting.

Table 5.

The performance of AI models for binary prediction of SA‐only and SA‐BC.

Items		Precision		Sensitivity		F1‐score		AUC	Accuracy	C_index
		SA	SA‐BC	SA	SA‐BC	SA	SA‐BC
Model	LR	0.890	0.690	0.880	0.700	0.880	0.690	0.866	0.832	0.789
	SVM	0.730	0.000	1.000	0.000	0.840	0.000	0.862	0.729	0.500
	LSVM	0.860	0.680	0.890	0.610	0.880	0.640	0.853	0.814	0.750
	DT	0.840	0.540	0.820	0.580	0.830	0.560	0.710	0.753	0.699
	RF	0.900	0.770	0.920	0.720	0.910	0.750	0.912	0.866	0.820
	XGBoost	0.910	0.800	0.930	0.760	0.920	0.780	0.925	0.883	0.844
Features	Top 4	0.920	0.770	0.920	0.770	0.920	0.770	0.909	0.876	0.844
	Top 7	0.910	0.780	0.920	0.760	0.920	0.770	0.925	0.876	0.840
	Top 7 + 1	0.910	0.800	0.930	0.760	0.920	0.780	0.925	0.883	0.844

Abbreviations: AUC. area under the ROC curve; BC, breast cancer; DT, decision tree; LR, logistic regression; LSVM, linear support vector machine; RF, random forest; ROC, receiver operating characteristic; SA, sclerosing adenosis; SVM, support vector machine; XGBoost, eXtreme Gradient boosting.

Table 6.

The performance of Al model for ternary prediction of SA‐only, SA‐DCIS, and SA‐IBC.

Items		Precision			Sensitivity			F1‐score			AUC	Accuracy	C‐index
		SA	SA‐DCIS	SA‐IBC	SA	SA‐DCIS	SA‐IBC	SA	SA‐DCIS	SA‐IBC
Model	LR	0.800	0.000	0.400	0.920	0.000	0.410	0.860	0.000	0.410	0.776	0.739	0.647
	SVM	0.730	0.000	0.000	1.000	0.000	0.000	0.840	0.000	0.000	0.833	0.729	0.500
	LSVM	0.770	0.000	0.380	0.940	0.000	0.280	0.850	0.000	0.330	0.794	0.729	0.595
	DT	0.880	0.260	0.470	0.810	0.330	0.540	0.840	0.290	0.510	0.693	0.715	0.754
	RF	0.880	0.410	0.630	0.930	0.390	0.480	0.910	0.400	0.540	0.863	0.801	0.789
	XGBoost	0.900	0.430	0.550	0.950	0.300	0.520	0.930	0.360	0.530	0.888	0.811	0.813
Features	Top 4	0.870	0.550	0.550	0.940	0.480	0.390	0.910	0.520	0.460	0.856	0.804	0.763
	Top 7	0.890	0.480	0.550	0.950	0.300	0.520	0.920	0.370	0.530	0.885	0.808	0.793
	Top 9	0.870	0.450	0.690	0.930	0.390	0.520	0.900	0.420	0.590	0.877	0.808	0.776
	Top 7 + 1	0.900	0.430	0.550	0.950	0.300	0.520	0.930	0.360	0.530	0.888	0.811	0.812

Abbreviations: AUC. area under the ROC curve; DCIS, ductal carcinoma in situ; DT, decision tree; IBC, invasive breast cancer; LR, logistic regression; LSVM, linear support vector machine; RF, random forest; ROC, receiver operating characteristic; SA, sclerosing adenosis; SVM, support vector machine; XGBoost, eXtreme Gradient boosting.

Figure 3.

(a) Calibration curves for each model in binary classification in the test cohort to predict SA and SA + BC. (b) Calibration curves for each model in binary classification in the temporal validation cohort to predict SA and SA + BC(N = 113). BC, breast cancer; SA, sclerosing adenosis.

Additionally, to validate the rationale behind feature selection, we compared the predictive performance of the XGBoost model incorporating the top 4, top 7 and top 7 + 1 features on the feature selection list (Supporting Information: Table 5) from both machine learning and the classical nomogram model, revealing that the integration of top 7 with the 1 feature from nomogram as the top 7 + 1 method had the best performance (Table 5).

3.5. Temporal Validation With Latest Data Set

To assess temporal generalizability, we evaluated all models in an independent external cohort. XGBoost again delivered the strongest performance across metrics: binary classification accuracy 0.8584, C‐index 0.8226, AUC 0.9038; ternary classification accuracy 0.7257, C‐index 0.7396, AUC 0.8119 (Table 7). As expected in temporal validation, all models exhibited some performance attenuation relative to the original test set; however, XGBoost showed the smallest degradation, indicating the most robust generalization. The relative ranking of methods remained stable, with tree‐based approaches (XGBoost, RF) outperforming linear models for both tasks, supporting the durability of our feature‐selection and modeling strategy.

Table 7.

Prediction performance of binary and ternary models in the temporal validation cohort.

Model	Binary			Ternary
	Accuracy	C_index	AUC	Accuracy	C_index	AUC
LR	0.7876	0.7226	0.8452	0.6814	0.5981	0.7658
SVM	0.8142	0.8223	0.8726	0.6460	0.5000	0.7967
LSVM	0.7965	0.7295	0.8966	0.6726	0.5530	0.8069
DT	0.6991	0.6824	0.6824	0.6372	0.7310	0.6385
RF	0.8142	0.7771	0.8848	0.6991	0.7402	0.7909
XGBoost	0.8584	0.8226	0.9038	0.7257	0.7396	0.8119

Abbreviations: AUC. area under the ROC curve; DT, decision tree; LR, logistic regression; LSVM, linear support vector machine; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine; XGBoost, eXtreme Gradient boosting.

Focusing on the clinically critical binary task (SA‐only vs. SA + BC) in the temporal cohort, calibration curves showed that XGBoost and RF achieved the closest alignment of reliability curves to the ideal diagonal, outperforming other benchmarks (Figure 3b). DCA further supported clinical utility, with XGBoost yielding higher net benefit across a wide range of decision thresholds compared with “treat‐all” and “treat‐none” strategies (Figure 2d).

3.6. Illustrated Explanation of Global, Local and Individualized Feature Contribution

Using SHAP, we visualized global, local, and individual feature contributions for both the binary and ternary models. Global importance was quantified by mean absolute SHAP values (Figures 4 and 5). Local feature effects across the cohort were depicted with SHAP summary (swarm) plots (Figures 6 and 7), illustrating how feature values shift predictions. Case‐level interpretability was provided via SHAP force plots, showing individualized feature contributions to predicted risk for representative patients (Figures 8 and 9). These complementary views enhance transparency and clinical interpretability of the models.

Figure 4.

The global contribution in SA and SA‐BC classification prediction. BC, breast cancer; MG, mammography; SA, sclerosing adenosis; US, ultrasound.

Figure 5.

The global contribution in SA & SA‐DCIS and SA‐IBC classification prediction. DCIS, ductal carcinoma in situ; IBC, invasive breast cancer; MG, mammography; SA, sclerosing adenosis; US, ultrasound.

Figure 6.

The local contribution in SA and SA‐BC classification prediction. BC, breast cancer; MG, mammography; SA, sclerosing adenosis; US, ultrasound

Figure 7.

The local contribution in SA & SA‐DCIS & SA‐IBC classification prediction. DCIS, ductal carcinoma in situ; IBC, invasive breast cancer; MG, mammography; SA, sclerosing adenosis; US, ultrasound.

Figure 8.

The SHAP force plot of SA and SA‐BC patients' cases. The first patient was correctly predicted as SA‐only with SHAP value of −3.44. The palpable mass (biopsy procedure = 0), the young age of 38 (age = 38), the US minimum diameter of 0.6 and a US BI‐RADS 4a (US BI‐RADS = 4.1) were all indicated benign. The second patient was correctly predicted as SA‐only with SHAP value of −1.51. The US maximal diameter of 2.4 (US Diameter Max = 2.4), US minimal mass diameter of 1.1 (US Diameter Min = 1.1) and a US BI‐RADS 4c (US BI‐RADS = 4.7) were all associated with increased risk of cancer, whereas age of 31 and the NPBL nature were linked to a decreased risk of BC. The other two patients were both correctly predicted as SA‐BC. BI‐RADS, breast imaging reporting and data system; BC, invasive breast cancer; SA, sclerosing adenosis; US, ultrasound.

Figure 9.

The SHAP force plot of SA & SA‐DCIS & SA‐IBC patients' cases. Each force diagram for ternary prediction of individual patient was actually a superposition of three force diagrams of binary prediction. The higher the SHAP value, the more accurate was the current classification. For the first two patients, red features indicated an increased probability of the patient being predicted as SA‐only. Similarly, the two SA‐DCIS patients and the two SA‐IBC patients were all correctly predicted. DCIS, ductal carcinoma in situ; IBC, invasive breast cancer; SA, sclerosing adenosis; SHAP, Shapley additive explanations; US, ultrasound.

4. Discussion

SA is a common proliferative breast lesion, present in 62.4% of biopsies with proliferative disease without atypia and 55.1% of biopsies with ADH, and as an isolated finding approximately doubles subsequent BC risk [6]. Clinically and radiologically, SA can closely mimic malignancy, leading to potential overdiagnosis; more critically, SA may coexist with carcinoma, underscoring the need to detect concurrent disease when core‐needle biopsy (CNB) or fine‐needle aspiration (FNA) indicates SA [11, 12, 13, 47, 48]. While the guideline recommends that SA alone is not an indication for surgical biopsy or excision [49], our findings suggest that when SA is diagnosed on CNB, surgical excision should be considered to exclude coexistent DCIS or IBC.

Existing diagnostic models largely compare SA‐only with BC‐only cohorts [50, 51]. For example, Liang et al. developed a six‐parameter nomogram (age, calcification, echogenic rim, vascularity distribution, tumor size, parenchymal thickness) that distinguishes SA from BC with strong performance [13]. However, this nomogram required that the SA and BC lesions have the same US BI‐RADS classification, and it failed to further distinguish simple SA from SA coexisted with DCIS or IBC. The challenge is evident even intraoperatively: among patients undergoing FFB, 19.4% (24/124) yielded indeterminate results (Supporting Information: Table 2). Notably, in the SA + DCIS subgroup, 18.9% (10/53) had uncertainty regarding invasive components, whereas the uncertainty rate in SA‐only cases was 0%, indicating that FFB, reported as SA reliably reflects SA in SA‐only patients (Supporting Information: Table 2). Given the divergent prognoses and treatment strategies for DCIS versus IBC, precise differentiation among SA‐only, SA + DCIS, and SA + IBC is essential. To our knowledge, this study is the first to develop and validate predictive models that differentiate SA‐only from SA + BC (binary) and further stratify SA + DCIS versus SA + IBC (ternary), integrating interpretable AI with clinically grounded nomograms. This framework addresses a key unmet need by enhancing preoperative risk stratification and informing surgical decision‐making when SA is identified on CNB.

Our integrated pipeline—spanning feature selection, classification, and model explanation—demonstrates a robust, AI‐driven approach for medical prediction, capable of mining large, heterogeneous datasets to uncover patterns not readily discernible to human observers [31, 39, 44, 52, 53, 54]. To ensure methodological rigor, we employed six complementary feature‐selection techniques and six classifiers for predictive modeling. We further leveraged nomograms, a widely used tool for individualized risk estimation that synthesizes multiple variables [55], to enhance model stability and clinical interpretability. Specifically, we augmented the ML‐derived feature set with key variables from both the binary and ternary nomograms and integrated these features into the classifiers, which yielded additional performance gains. This hybrid, clinically informed ML strategy balances predictive accuracy with transparency, facilitating translation into decision support.

We selected eight practical predictors for the streamlined model: age, US BI‐RADS, minimum and maximum US diameters, US margin (boundary), MG micro‐calcifications, MG density, and biopsy procedure (Tables 5 and 6 and Figures 2 and 3). BC risk increases with age, and age is commonly included in radiomics and deep‐learning models distinguishing SA from BC [13, 56]. Although US BI‐RADS is partially operator‐dependent, it remains an independent discriminator of malignancy [28, 57], and its objectivity can be enhanced by computer‐aided texture analysis [50]. Lesion size metrics are established predictors in nomograms [58, 59]; notably, our study specifically highlights the US minimum diameter, which correlated with less invasive behavior and better compliance. As a key determinant of three‐dimensional lesion shape, a smaller minimum diameter relative to the maximum suggests more discoid morphology (e.g., min≈0.5 × max) [60]. Malignancies more often show ill‐defined or spiculated margins, whereas SA tends to present with circumscribed margins [13]; boundary‐aware deep learning has further improved ultrasound lesion delineation [61, 62]. The “biopsy procedure” feature captures clinical presentation and workflow (e.g., direct excisional biopsy of palpable masses and excisional biopsy of non‐palpable breast lesions requiring image‐guided localization), with prior evidence that non‐palpable breast lesions (NPBLs) are associated with SA [50, 63]. MG microcalcifications and density remain foundational imaging cues for risk stratification; dense breasts both elevate BC risk and obscure detection [64]. MG density was also retained based on its consistent selection in the nomogram, reinforcing its clinical relevance (Supporting Information: Table 5).

XGBoost, a gradient‐boosting algorithm, is well suited to large datasets and offers stable feature selection and training. In our study, it delivered the strongest performance among all classifiers: for the binary task, AUC 0.925, accuracy 0.883, and C‐index 0.844; for the ternary task, AUC 0.888, accuracy 0.811, and C‐index 0.813. This robust discrimination between SA‐only and SA + BC provides objective support for clinical decision‐making, especially given that approximately 20% of SA + BC cases had indeterminate intraoperative FFB results (Supporting Information: Table 2).

To assess durability over time, we performed temporal validation in a recent prospective cohort. As expected with potential shifts in case mix, performance showed modest, acceptable attenuation; however, the framework consistently outperformed benchmark models, with notably strong robustness in the binary classification task. The larger relative decline observed in ternary performance likely reflects the limited sample size of the temporal cohort (n = 113), which constrains reliable learning and evaluation of three‐class discrimination. Overall, these results support the sustained clinical utility of the XGBoost‐based framework.

Lack of interpretability can render ML models “black boxes,” obscuring how predictions are generated. SHAP‐based explainability addresses this by quantifying and visualizing each feature's contribution and has been widely adopted across medical diagnosis, decision support, and treatment planning [53, 65, 66]. In this study, we provided multi‐level interpretability: global importance via mean absolute SHAP values, cohort‐level local effects via summary plots, and case‐level explanations via force plots. US BI‐RADS, age, and minimum US diameter consistently showed the highest overall contributions (Figures 4 and 5). Individual force plots for both binary and ternary tasks illustrated how patient‐specific features acted as “forces” (SHAP values) shifting predictions toward or away from malignancy (Figures 8 and 9).

There were several limitations in this study. First, as a retrospective study, the data were accrued over an extended period, and although temporal validation supported robustness, natural shifts in patient characteristics may affect long‐term performance. Prospective studies with periodic recalibration or dynamic updating are warranted. Secondly, despite our comprehensive feature selection strategy, integrating more refined data, such as radiomics, in the future may capture more subtle phenotypic patterns and further enhance performance. Thirdly, the high morphological similarity and feature overlap between SA‐DCIS and SA‐IBC posed a significant challenge to the machine learning models, potentially leading to lower sensitivity for SA‐DCIS detection. Improving the sensitivity for SA‐DCIS, possibly through cost‐sensitive learning or advanced feature extraction, is a critical goal for future work. Fourthly, all the patients did not undergo CNB and the case number or percentage of SA + BC with a SA‐only CNB was unclear, therefore the actual under‐estimate rate of SA + BC mistakenly as SA‐only was unavailable. Finally, among the 537 SA‐only patients, the actual case number or percentage of uncertain intra‐operative FFB was also unclear.

5. Conclusion

By integrating a traditional nomogram with an AI framework, we combined eight clinically actionable features—age, US BI‐RADS, maximum and minimum US diameters, US margin (boundary), biopsy procedure, MG density, and MG microcalcifications—into an XGBoost model that achieved strong performance for both binary (SA‐only vs. SA + BC) and ternary (SA‐only vs. SA + DCIS vs. SA + IBC) classification. Given the risk of coexistent carcinoma, particularly DCIS or IBC, when SA is identified on core‐needle biopsy, our findings support considering surgical excision to definitively exclude concurrent malignancy.

Author Contributions

Yang Qu: methodology, investigation, writing – original draft, resources, conceptualization. Jie Lian: writing – original draft, methodology, resources, conceptualization. Tianli Liu: methodology, software, formal analysis, conceptualization. Ying Xu: software, methodology, conceptualization, writing – review and editing. Zhe Wang: resources, investigation, validation. Futian Weng: methodology, software, writing – review and editing. Jiahui Zhang: writing – review and editing, formal analysis. Xu Yang: formal analysis, investigation, data curation. Jing Qin: resources, methodology. Ming Wang: data curation, software. Wen Xu: resources, methodology. Wenbo Li: writing – original draft, resources. Lingyan Kong: conceptualization, data curation, resources, visualization. Xinyu Ren: resources, visualization. Qiang Sun: writing – review and editing, conceptualization, supervision. Bo Pan: supervision, conceptualization, funding acquisition, writing – review and editing, project administration. Yidong Zhou: conceptualization, writing – review and editing, supervision, project administration. Yan Xu: resources, supervision, conceptualization, project administration.

Ethics Statement

Ethical approval was granted by the Ethics Committee of Peking Union Medical College Hospital, Chinese Academy of Medical Sciences(approval number: K3610).

Consent

Written informed consent was waived due to the anonymized nature of the data.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: Construction and validation of nomogram models to predict SA and SA+BC. Figure S2: Construction and validation of nomogram models to predict SA, SA+DCIS, SA+IBC.

Table S1: Rational, equations and references in feature selection, model establishment and performance evaluation and explanatory analysis in AI deep learning. Table S2: The pathological characteristics of the SA‐only and SA+BC patients. Table S3: Parameters include area under curve (AUC), accuracy and C‐index to evaluate the accuracy and discrimination of binary and ternary categorized nomogram models. Table S4: Top five clinical and imaging variables selected by various feature selection methods. Table S5: Feature selection summary of how many times each variable was selected. Table S6: Detailed search space and final optimization parameters.

Qu Y., Lian J., Liu T., et al., “Binary and Ternary Classification Prediction for Breast Cancer and Breast Sclerosing Adenosis With Interpretable Artificial Intelligence From Clinical and Imaging Features: A Retrospective, Diagnostic Accuracy Cohort Study,” Cancer Innovation 5 (2026): e70049. 10.1002/cai2.70049.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. The data sets analyzed during the current study are available from the corresponding author on reasonable request.