AI Model Predicts Axillary Response Using Baseline Mammography

Refining Axillary Management After Neoadjuvant Therapy

Neoadjuvant therapy, which involves systemic treatment administered before surgical resection, has become a standard of care for locally advanced breast cancer, offering the opportunity to downstage tumors and evaluate in vivo treatment response [1, 2]. While systemic regimens including dual human epidermal growth factor receptor 2 (HER2) blockade and immunotherapy have significantly improved pathological complete response rates, with some triple-negative protocols achieving an odds ratio of 3.95 (95% CrI 1.81 to 9.44) compared to standard chemotherapy [3], managing the axilla post-treatment remains a clinical challenge [4]. Current surgical standards often necessitate invasive axillary lymph node dissection for patients who were node-positive at diagnosis, despite the risk of long-term complications such as lymphedema and psychological distress, which can affect up to 78.5% of patients after neoadjuvant therapy [5]. Although sentinel lymph node biopsy, a procedure to identify and remove only the primary draining nodes, is an option for some, its pooled false-negative rate of 0.14 (95% CI 0.11 to 0.17) remains a concern in the post-neoadjuvant setting [6]. A new study now evaluates whether artificial intelligence can leverage baseline imaging to provide more precise guidance on axillary status.

Architecture of the Deep Learning Model

The researchers developed a deep learning-based artificial intelligence model designed to analyze baseline digital mammography images to predict axillary lymph node status following neoadjuvant therapy. This multicenter study utilized images and clinical data from 956 patients with invasive non-specific breast cancer across three separate medical institutions. Every patient in this cohort presented with confirmed positive axillary lymph node metastasis at baseline, establishing a consistent starting point for evaluating how well the model could identify nodal conversion, which is the transition from node-positive to node-negative status after systemic treatment. To optimize predictive accuracy, the authors compared four distinct image cropping methods and five different backbone networks (the underlying architectural frameworks used for feature extraction and classification). The analysis determined that a fixed 5 cm image clipping method provided the most effective input. For primary image processing, the researchers identified the Swin Transformer V2 as the most effective backbone feature extraction network, which is a specific type of neural network architecture that uses a hierarchical approach to process visual data and capture complex patterns within the image. By refining these technical parameters, the model was better equipped to identify subtle radiographic features in the primary lesion that correlate with axillary response. The integration of these architectural choices allowed the model to process both primary and auxiliary region images alongside clinical features. In the training set, this comprehensive approach yielded an area under the receiver operating characteristic curve (AUC) of 0.823 (95% CI: 0.797 to 0.846, p < 0.001). Similar performance was maintained across the internal validation set, which showed an AUC of 0.774 (95% CI: 0.722 to 0.818, p < 0.001), and the external test set, which achieved an AUC of 0.756 (95% CI: 0.700 to 0.805, p = 0.013). By identifying patients who achieve nodal conversion, this tool provides clinical decision support that may help physicians avoid performing unnecessary and morbid axillary dissections in patients who have already responded to therapy.

Predictive Accuracy Across Multi-Center Datasets

The researchers observed that the predictive capability of the model improved significantly when they moved beyond analyzing the primary tumor alone. By adding a pre-training model (a technique where the artificial intelligence is first trained on a large, general dataset to recognize basic visual patterns before being fine-tuned for specific medical tasks) and integrating clinical features with the primary lesion input, the model achieved an AUC of 0.823 in the training set (95% CI: 0.797 to 0.846, p < 0.001). This statistical measure, where an AUC of 1.0 represents perfect prediction, indicates a high level of accuracy in distinguishing between patients who achieved nodal conversion and those with residual axillary disease. The model maintained consistent performance across different cohorts, reaching an AUC of 0.774 in the internal validation set (95% CI: 0.722 to 0.818, p < 0.001) and an AUC of 0.778 in the internal test set (95% CI: 0.739 to 0.813, p = 0.034). To ensure the tool remained reliable across different clinical environments and imaging equipment, the authors evaluated its performance using an external test set from a separate medical center. In this external cohort, the model achieved an AUC of 0.756 (95% CI: 0.700 to 0.805, p = 0.013), suggesting that the deep learning features identified on baseline mammography are generalizable across different patient populations. The most robust results occurred when the researchers utilized a comprehensive input strategy that combined images of both the primary tumor and the auxiliary regions with the patient's clinical data. With this multi-faceted approach, the AUC values reached above 0.8 across all four datasets, including the training, internal validation, internal test, and external test sets. For the practicing clinician, this level of predictive accuracy provides a data-driven method to identify candidates for less invasive axillary surgery after neoadjuvant therapy, potentially reducing the long-term morbidity associated with full nodal dissections.

Clinical Utility in Reducing Surgical Overtreatment

The clinical management of the axilla following neoadjuvant therapy remains a significant challenge due to the lack of reliable methods to accurately judge the status of axillary lymph nodes. Currently, the inability to precisely identify which patients have achieved a pathologic complete response in the lymph nodes often leads to overtreatment, as clinicians may default to more invasive surgical interventions to ensure oncologic safety. This study addresses this diagnostic gap by utilizing an artificial intelligence model based on baseline digital mammography images taken before the start of neoadjuvant therapy. By analyzing these initial scans of 956 patients with invasive non-specific breast cancer and confirmed nodal metastasis, the researchers demonstrated that deep learning can extract predictive features of nodal conversion that are not visible to the human eye during routine radiological review. The researchers conclude that the model provides critical decision support to avoid excessive surgery by accurately judging axillary lymph node status after systemic treatment. When the model integrated primary and auxiliary region images with clinical features, it achieved an AUC above 0.8 across the training, internal validation, internal test, and external test sets, providing a robust statistical foundation for surgical planning. For the practicing clinician, this means that baseline imaging may hold the key to identifying which patients are likely to have negative nodes following therapy, potentially allowing for the omission of axillary lymph node dissection. By providing a more accurate assessment of the axillary response, this artificial intelligence tool may help in the clinical selection of more beneficial, less invasive treatment modalities, directly mitigating the long-term morbidity associated with surgical overtreatment in breast cancer care.

References

1. Qari AS, Mowais AH, Alharbi SM, et al. Adjuvant and Neoadjuvant Therapy for Breast Cancer: A Systematic Review.. European journal of breast health. 2024. doi:10.4274/ejbh.galenos.2024.2023-12-16

2. Lorenzo R, Meani F, Longhitano C, et al. The optimal timing between neoadjuvant therapy and surgery of breast cancer: A brief systematic review of the literature.. Critical reviews in oncology/hematology. 2023. doi:10.1016/j.critrevonc.2023.103921

3. Lin Y, Gao H, Yang X, et al. Neoadjuvant therapy in triple-negative breast cancer: A systematic review and network meta-analysis.. Breast (Edinburgh, Scotland). 2022. doi:10.1016/j.breast.2022.08.006

4. Villacampa G, Matikas A, Oliveira M, Prat A, Pascual T, Papakonstantinou A. Landscape of neoadjuvant therapy in HER2-positive breast cancer: a systematic review and network meta-analysis.. European journal of cancer (Oxford, England : 1990). 2023. doi:10.1016/j.ejca.2023.03.042

5. Omari M, Amaadour L, Asri AE, et al. Psychological distress and coping strategies in breast cancer patients under neoadjuvant therapy: A systematic review.. Women's health (London, England). 2024. doi:10.1177/17455057241276232

6. Vázquez JC, Piñero A, Castro FJD, et al. The value of sentinel lymph-node biopsy in women with node-positive breast cancer at diagnosis and node-negative tumour after neoadjuvant therapy: a systematic review.. Clinical & translational oncology : official publication of the Federation of Spanish Oncology Societies and of the National Cancer Institute of Mexico. 2023. doi:10.1007/s12094-022-02953-1