Purpose: Accurate diagnosis of axillary lymph node metastases (ALNMs) is critical for breast cancer staging and management, particularly since the advent of neoadjuvant chemotherapy. Ultrasound is the most widely used pre-operative modality to detect ALNMs, yet sensitivity and specificity has been reported as low as 61% and 85%, respectively. While CT chest studies are often performed for breast cancer staging in locally advanced breast cancer (LABC), they have demonstrated limited accuracy for the detection of ALNMs based on conventional features of lymphadenopathy. This study aims to determine whether CT radiomic analysis improves detection of ALNMs compared to conventional features in patients with (LABC).
Materials and Methods: Retrospective chart review was performed on patients referred to a tertiary LABC clinic between 2016-2019 who demonstrated a positive ALNM upon ultrasound-guided fine needle aspiration, underwent subsequent CT chest imaging and had no prior history of cancer. Pre-biopsy ultrasound imaging was used to identify biopsy-proven lymph nodes on CT chest imaging for each patient. A contralateral lymph node was used as a negative control for each patient. Both positive and negative lymph nodes were analyzed for conventional features of lymphadenopathy (long axis ≥ 1 cm, short axis ≥ 1 cm, long-to-short ratio ≤ 2, thickened cortex, absent hilum, matted nodes) and for 111 radiomic features extracted using PyRadiomics software following manual 3D segmentation.
Results: 75 patients (150 lymph nodes) met inclusion/exclusion criteria and were assessed for both conventional and radiomic features. The strongest conventional feature taking into account both sensitivity and specificity was an absence of fatty hilum, with a sensitivity and specificity of 75% for ALNMs. Of the 111 extracted radiomic features, 24 features demonstrated area under the curve (AUC) values greater than 0.90. These features all demonstrated combined sensitivities and specificities higher than the strongest conventional feature, with energy demonstrating an AUC of 0.96, sensitivity of 95% and specificity of 81%. 94% of extracted radiomics features demonstrated good intraclass correlation coefficients (ICCs) (>0.70) and 60% of features demonstrated excellent ICCs (>0.90).
Conclusion: Radiomic analysis improves detection of ALNMs in breast cancer compared to conventional CT features. Implementation of modelling combining clinical and radiomic features may further improve diagnostic accuracy.
Clinical Relevance Statement: Development of an accurate and non-invasive tool to identify nodal metastases in breast cancer may lead to more accurate staging and optimal neoadjuvant treatment decisions.
