Medical Imaging & AI in Oncology

Introduction

The axillary lymph node (ALN) status represents one of the most important independent prognostic factors for predicting disease-free survival and overall survival in patients with breast cancer (BC), because it reflects the risk of distant recurrence and death after loco-regional treatment [1]. The 5-year survival rate drops from 98.6% to 84.4% when ALN positivity is present at diagnosis [2]. Therefore, accurate assessment of ALN metastases is critical for the clinical decision-making process. Currently, the gold standard for axillary staging in patients with BC is the histological examination of sentinel lymph-node biopsy (SLNB) and/or axillary lymph-node dissection (ALND). Unfortunately, both procedures are invasive and have a high incidence of complications, such as lymphedema, upper limb neuropathy, and limited movement of the shoulder [3,4]. In current clinical practice, ALN status can be preoperatively assessed using different imaging techniques, such as US, magnetic resonance imaging (MRI) and 18F-FDG Positron Emission Tomography/Computed Tomography (PET/CT) [5]. Among these, ultrasound (US) plays a leading role, being widely available, easy to perform, cost effective, and able to guide biopsy. This technique has high sensitivity (87%) in assessing ALN status but still suffers from low and variable specificity (53–97%) due to high operator-dependence, while providing qualitative information [6]. The clinical implications of US assessment of ALN status have become even more relevant in light of the evidence from the SOUND (Sentinel Node vs. Observation after Axillary Ultrasound) randomized clinical trial. Indeed, it demonstrated that omitting axillary surgery was noninferior to SLNB in patients with BC up to 2 cm and a negative axilla on US [7,8]. Therefore, the development of a noninvasive ALN staging method to accurately assess BC clinical stage and select tailored treatment options for patients is an urgent need. In this scenario, a relatively new technique, breast shear-wave elastography (SWE), has been introduced to enhance the US diagnostic accuracy in characterizing breast lesions and, consequently, improve patient management [9,10]. It provides qualitative data, shown as a semitransparent color-coded image, and quantitative data, expressed as shear wave velocity (m/sec) or elasticity (kPa) measured within a specified region of interest (RO [11]. The potential enhancing role of SWE in predicting ALN status in patients with BC has not been investigated to date [12,13]. Over the last few years, the possibility of extracting quantitative data from medical images, so-called radiomics, has opened new research perspectives for detecting tumor features invisible to the human eye and possibly correlated with cancer heterogeneity and behavior, using artificial intelligence (AI) software [14,15]. Regarding the potential applications of radiomics and AI in the diagnosis and management of BC, several investigations have used US, digital breast mammography, or MRI with different aims, including differential diagnosis of breast lesions, prediction of BC molecular subtypes, and assessment of ALN status [16-18]. In detail, regarding this latter task, most machine learning (ML) models were built using quantitative parameters extracted from primary tumor lesions on MR image [19,20].

Performance Comparison: ML vs. Radiologist

Conclusion

The built ML model included both US B-mode and SWE-derived radiomics features, showing a moderate performance in predicting ALN status, inferior, even if not significantly, to that of an expert radiologist. Our findings support the use of SWE-derived radiomics features in the ML radiomics pipeline and, given the important role of US in defining ALN, additional investigations on larger multicentric datasets are encouraged to further reinforce the current evidence.