patient
Informed consent for participation was exempt by the Institutional Review Board (IRB Approval Number: 2023ER131-1) of the affiliated hospitals of North Sichuan Medical University, in accordance with national regulations. This study involved a retrospective analysis of 344 patients with pathologically confirmed BC assessed in our hospital from June 2021 to December 2023.
Inclusion criteria: (1) Pathologically confirmed IBC surgery of puncture. (2) LNM status was confirmed by surgery or puncture pathology. (3) Complete pathological and immunohistochemical results. (4) No prior biopsy, surgery, chemotherapy, or radiation therapy prior to MRI examination. (5) DCE and DWI scans were performed prior to treatment.
Exclusion criteria: (1) Unable to meet the analysis requirements in image quality. (2) Incomplete clinical data.
Based on these criteria, 183 IBC patients were enrolled in this study. This includes 107 people including ALNM and 76 people from the non-ALNM group (Figure 1). The median age for these patients was 51 years (range: 24–84).
Patient selection strategies. DWI: Diffusion-weighted imaging. DCE: Dynamic contrast enhancement.
Acquiring images
All imaging data from this study were collected using a United Imaging Healthcare 3T MRI scanner (UMR790 3.0T; Shanghai, China) with a 10-channel breast-specific phased array coil. The patient had her head facing forward and her chest was positioned naturally within the chest coil, and the center of the coil was aligned with the positioning center and nipple lines on both sides. Bilateral breast and x-fossa scans were performed. Axial T1 weighted imaging multiphase DCE MRI (TE: 2.17 ms, TR: 5.15 ms, layer thickness 1.0 mm, field of view 340 mm x 340 mm, matrix 336 x 100, 128 layers) and Axial DWI (TE: 66 ms, TR: 4000 mm x 190 mm, field thickness 4.0 mm, field 190 mm, × 100, 24 layers2respectively). One-phase montage scans were performed prior to contrast injection. GD-DTPA (Jadixian, Jiangsu Hengareutical Co., Ltd.) was then injected via the central vein of the cube vein in a high-pressure syringe at a dose of 0.2 mmol/kg and a flow rate of 2.5 ml/s, followed by a 20 ml layer at this same rate. After this injection, eight consecutive scan phases (69 s/phase) were started due to an overall scan time of 9 min 36 s.
Clinical data
Patients' age and menopause status were collected from the Clinical Medical Records System. It was defined as the presence of nuclear staining in ≥1% of tumor cells according to the American College of Clinical Oncology (ASCO) and the American School of Pathology (CAP), guidelines for immunohistochemical detection in breast cancer issued by the estrogen receptor (ER) and progesterone receptor (PR) positivity. The negativity of ER and PR was defined as nuclear staining of less than 1% of tumor cells in the presence of positive internal controlstwenty five. HER-2 negativity (HER-2-) was defined as an immunohistochemistry (IHC) score of 0 or an IHC score of 1+ for Insitu hybridization (ISH) that does not exhibit HER-2 gene amplification. HER-2 positive (HER-2+) is defined as an IHC score of 3+, or IHC score 2+, and ISH indicates HER-2 gene amplification26. High Ki-67 expression was defined as above 14%, and low expression was defined as below 14%27.
Based on the expression status of ER, PR, HER-2, and KI-67, all cases were divided into four subtypes: Luminal A (ER/PR positive, HER-2 negative, low Ki-67 expression), Luminal B (ER/PR positive with either high KI-67 expression or HER-2 positive), HER-2 overexpression (ER-2 positive), ER-2 positive), and ER 2.27,28.
Radiomics Features Extraction
Regions of interest (ROI) were drawn using a 3D Slicer (V 5.0.3, https://www.slicer.org/), followed by feature extraction (Figs. 2 and 3). Two imaging physicians with a diagnostic breast MRI experience of over 5 years were independently responsible for selecting the amount of interest by defining breast tumor ROI individually in a layer-by-layer manner using images of the second contrast phase of the DCE sequence (B = 800 S/mm image).2) DWI images while blinded by patient LNM status. In patients with multiventricular or multifocal tumors, only the lesions with the largest diameter were analyzed. Using a Python-based approach of 3D slicers using the “Pyradiomics” package, we extracted seven general feature groups from these ROIs, resulting in a total of 1,223 features.
A 41-year-old woman with a special type of invasive cancer in the right breast, luminal A subtype without lymph node metastasis. (a, b) Original DWI images and a schematic diagram of an overview of the area of interest. (c, d) Original DCE images and a schematic diagram of the area of interest.
A 54-year-old female patient with a special type of invasive cancer in the left breast, a HER-2 overexpressing subtype, with lymph node metastasis. (a, b) Original DWI images and a schematic diagram of an overview of the area of interest. (c, d) Original DCE images and a schematic diagram of the area of interest. NOTE: DWI is diffusion-weighted imaging, and DCE is dynamic contrast enhanced imaging.
Evaluation of interobserver contract tests
Approximately one-third of the DCE and DWI images were randomly selected from all patients included in this study, and both participating physicians drew ROIs for the lesions in these images followed by extraction of the corresponding radioactive features. We use interclass correlation coefficient (ICC) values to measure interobserver contracts by evaluating the extracted data, but excludes the ICC <0.75 feature and ICC ≥0.75 feature. Ultimately, this preserved 1,138 and 923 features in the DCE and DWI sequences, respectively.
Functional selection and model development
Features selection, model development, and model testing were performed in Python (v 3.9.12). Following Z-score normalization, features with variance >0.8 were retained via the variance threshold (VT) method. The selectkbest and the minimum absolute contraction and selection operator (Lasso) methods were employed for function selection. In total, eight features were screened in DWI sequences, including four Gy level size zone matrices (GLSZM), two Gy level-dependent matrix (GLDM), one Gy level run length matrix (GLRLM), and one primary function. Similarly, eight features were screened: two GLRLM, 2 GLDM, 1 GLSZM, 1 GY level cooccurrence matrix (GLCM), one first order, and type 1 function. A total of eight functions were also screened for a total sequence, including one shape, 2 GLRLM, 2 GLDM, and 3 GLCM functions.
The Random Forest method was used to construct three models using these image features, including DCE, DWI, and combined sequence models. Each model was developed from the eight features mentioned above. ROC and decision curves were generated for each model, and SHAP values were calculated for each function and visualized to provide greater interpretability for these models.
The 183 patients included in this study were randomized at a 7:3 ratio into a training cohort of 128 patients (75 ALNM, 53 non-ALNM) and a laboratory cohort containing 55 patients (32 ALNM, 23 non-ALNM). Model training was performed in the training cohort and then validated in the test cohort.
Statistical analysis
SPSS 26.0 (IBM, USA) and Python were used for all statistical analyses. Continuous data were reported as mean ± SD for successful distribution and compared with independent sample t-tests, whereas skewed data were reported as median (Q25, Q75) and compared with Mann-Whitney U tests. Count data were reported as case numbers and compared with χ tests. ICC values were used to assess extracted data to measure interobserver agreement. Values from each model as a tool to predict ALNM in IBC patients were assessed based on AUC, sensitivity, specificity, accuracy, and F1 values, along with plots of the ROC curves. Differences in AUC values between models were compared with the Delong test. The decision curve was used to assess net clinical benefits. both sides p<0.05 was considered important.
