Axillary lymph node metastasis (ALNM) is one of the most important prognostic factors determining survival in breast cancer.1 The status of the axilla determines the need for axillary lymph node dissection (ALND), axillary radiotherapy, and neoadjuvant or adjuvant chemotherapy.2 Accurate determination of the axillary status before treatment is critical for defining individualized treatment options.3 Age, tumor size, tumor quadrant, multifocality, histological grade, pathological type, receptor status, and molecular subtype are associated with ALNM.4–9

Radiomic analysis aims to contribute to diagnosis, treatment, and follow-up by extracting specific quantitative information about diseases that the human eye cannot perceive from medical images. With radiomic analysis, the aim is to maximize the information obtained from images by obtaining quantitative data about signal intensities and the spatial distribution of interpixel relationships.10

Recently, radiomics has attracted considerable attention in the medical field, especially in oncology. Successful results have been achieved in the diagnosis, treatment, and classification of breast cancer.11 Also, promising results were acquired in the differentiation of malignant and benign breast masses, in the estimation of the grade, receptor status, and subtypes of malignant tumors using radiomic features extracted from magnetic resonance imaging (MRI). This method is also promising in the prediction of neoadjuvant chemotherapy response in breast cancer. In studies performed with MRI, it was found that radiomics successfully helped even in the prediction of breast cancer recurrence .12–16

Radiomic features of axillary lymph nodes from T2-weighted (T2W) MR images were not successful in predicting ALNM.17 However, ALNM could be predicted with radiomic features obtained from T2W, diffusion-weighted (DW), and contrast-enhanced T1-weighted (T1+C) images of the primary tumor.18–24

Our aim in this study was to investigate the performance of radiomic features obtained from short tau inversion recovery (STIR) sequences of the primary tumor in predicting axillary metastasis.

In this study, 421 patients were evaluated retrospectively. Overall, 106 patients who did not undergo MRI before treatment, 48 patients whose SLNB or ALND data were not available, 52 patients who received neoadjuvant chemotherapy, 6 patients with recurrent disease, 3 patients with concurrent malignancy, and 41 patients with artifacts were excluded from the study. Thus, 165 patients with a mean age of 51.12 ± 11 (range, 28–82) were included in the study. While 92 (55.76 %) patients had axillary metastases, 73 (44.24 %) did not have axillary metastases. The mean lesion size was 24.78 ± 15 (6–120) mm.

Altogether, 667 features with ICC values above 0.7 were evaluated with ML. Wavelet filtered texture features, maximum 3D diameter, skewness, kurtosis, and maximum signal features showed a high correlation with ALNM. The features selected by the ML algorithms and their importance are presented in Figure 2.

The accuracy and AUC of the ML algorithms on the training set ranged from 57% to 86% and 0.50 to 0.95, respectively (Table 1). Among the ML algorithms, the best model was the CatBoost classifier (AUC, 0.95; accuracy, 84%). The receiver operating characteristic curve showing the success of the CatBoost classifier in predicting ALNM is presented in Figure 3.

The CatBoost classifier was evaluated on the test set. After tuning, the AUC, accuracy, recall, precision, and F1 score were 0.92, 84%, 89%, 85%, and 86%, respectively. The CatBoost classifier model had a sensitivity of 92.9% and a specificity of 86.4% in detecting ALNM. The classification report and confusion matrix showing the performance of the CatBoost model are presented in Figure 4.

Recently, less invasive treatment approaches for the axilla have been accepted. SLNB is preferred instead of ALND in early-stage breast cancer. Determining the condition of the axilla in the preoperative period is important in the development of surgical plans. For individualized and minimally invasive treatment options, it is important to determine the condition of the axilla before treatment.25 Among imaging methods, ultrasound (US) is the primary tool in the evaluation of the axilla. However, approximately 15% to 20% of patients with negative US findings have metastases in the SLNB. In mammography, 50% of level 1 axillary lymph nodes can be visualized, and levels 2 and 3 cannot be evaluated.26 It has been reported that MRI has similar sensitivity to US but less specificity in detecting nodal metastases.27

In previous studies, ALNM could be predicted by the radiomic features of the primary tumor in MRI. Yu et al. successfully predicted axillary lymph node status on T1+C, T2W, and DW MRI with a multiomic signature including radiomic features and clinicopathologic features obtained from lymph nodes and the primary tumor.19 Qiu et al. were successful in predicting ALNM using clinicopathological features, morphological features of lymph nodes from MR images, and radiomic features of the primary mass. Radiomic features obtained from DWI, T2W, and T1+C images of the primary mass were successful in predicting ALNM with an AUC of 0.806.20 Wang et al. were able to predict axillary metastases with an AUC of more than 0.80 with the radiomic and deep learning features obtained from dynamic contrast-enhanced (DCE) MRI.21 Chen et al. were able to predict ALNM with a nomogram created from the radiomic and clinicopathological features of the primary tumor obtained from DW and DCE MRI (AUC in the training and test sets, 0.80 and 0.71, respectively).22 Using DCE MRI, Liu et al. predicted ALNM with a tumoral and peritumoral radiomic signature (AUCs in the training and test sets were 0.872 and 0.863, respectively).23 Cui et al. predicted axillary lymph node status with radiomic features obtained from second postcontrast images on MRI with an AUC of 0.86 and an accuracy of 89%.24 However, none of these studies presented the performance of T2W images alone in predicting ALNM.

Dong et al. were successful in predicting sentinel lymph node metastasis with radiomic features obtained from fat-suppressed T2W and DW MRI images. With T2W alone, AUCs were 0.847 in the training set and 0.770 in the validation set. With DW MRI, they obtained AUCs of 0.847 in the training set and 0.787 in the validation set. When they combined the features obtained from T2W and DW images, the AUCs were 0.863 in the training set and 0.805 in the validation set.18

In our study, we predicted ALNM with an AUC of 0.92 and 84% accuracy in the test set with the radiomic features obtained from STIR sequences by modeling using machine learning analysis. The ML models used, except for Naive Bayes, Dummy Classifier, SVM - Linear Kernel, and Ridge Classifier, showed successful performance with an AUC of at least 0.73 and 74% accuracy. The CatBoost Classifier, Extra Trees Classifier, Light Gradient Boosting Machine, Gradient Boosting Classifier, and Extreme Gradient Boosting had AUCs above 0.93 and accuracy values above 86%. Among the models, the CatBoost Classifier, Extra Trees Classifier, Light Gradient Boosting, Gradient Boosting Classifier, and Random Forest Classifier had the highest performance, with AUCs of 0.95 and accuracy values of 84% to 86%. Among these models, the CatBoost Classifier had the highest recall and precision values, at 93% and 82%, respectively.

Radiomic features derived from T2W sequences detected ALNM with an AUC of up to 0.85. When DWI and DCE were combined with T2W, the AUC increased to 0.86.18 In our study, only STIR sequences reached an AUC of 0.92. Since STIR sequences are used in routine breast MRI in some centers, they may contribute to the prediction of ALNM. To our knowledge, STIR radiomics has not been used in studies investigating ALNM. STIR sequences stand out compared with other fat suppression techniques by providing more uniform fat suppression without being affected by magnetic field inhomogeneity. Although the signal-to-noise ratio is poor, it is useful because it includes both T1 and T2 contrast.

There is a correlation between tumor size and lymph node metastasis.9 In our study, the maximum diameter of the tumor and axillary metastases showed a high correlation. Kurtosis and skewness features, which evaluate intralesional homogeneity, were also correlated with ALNM.

Our study has some limitations. Its retrospective nature, small number of patients, and choosing the largest lesion in patients with more than one lesion are among the limitations of our study. Another limitation is that we did not divide the patients according to the number of metastatic lymph nodes.

In conclusion, radiomic features obtained from the primary tumor on STIR sequences have the potential to predict ALNM. STIR sequences are noninvasive and are currently used as a routine component of breast MRI in some centers. However, some lesions cannot be segmented because they have a signal intensity close to that of the parenchyma in STIR sequences.

Ethics committee approval was obtained from our institution for this retrospective study (Approval No. 202351425036).