BI-RADS assessment of solid breast lesions based on quantitative ultrasound and machine learning

Due to promising results of shear wave ultrasound (US) elastography (SWE) and homodyned-K (HK) features in assessing solid breast lesions prior to biopsy, there is a high interest in developing a multi-parametric approach in this context. The goal was to assess the added value of quantitative US (QUS) parameters to SWE and BI-RADS category, based on random forests classifier. US evaluation of breast lesions was performed as per standard clinical practice. For women who gave written consent, 3 SWE images were obtained with an Aixplorer US system (Supersonic Imagine, Aix-en-Provence, France) using a SL15-4 probe. Then, a 1-second cine-loop of radiofrequency images was acquired with the same system and probe, using a different acquisition mode. Then, the radiologist performed percutaneous biopsy as per standard procedure. Breast pathologists performed histopathology analyses. Lesions were anonymized; data analysis was performed on a workstation. The maximum elasticity Emax on each of three regions-of-interest (ROI) - supra-, intra-, and infratumoral - was computed directly from acquired SWE images, for a total of 3 features. For each of the three ROIs on B-mode images, the following HK parametric maps were considered based on previous work: 1) the infratumoral total signal power normalized by the maximal intensity in the ROI; 2) the intratumoral reciprocal of the scattering clustering parameter; 3) the supratumoral coherent-to-diffuse signal ratio; and 4) the supratumoral diffuse-to-total signal power ratio. Eight features were extracted from these maps. The total attenuation coefficient slope, assessed with the spectral fit method, was also computed in the intra- and infratumoral ROIs, for a total of 13 features. One hundred and three women with suspicious solid breast lesions (BI-RADS categories 4-5) were enrolled. When considering all types of features for these 103 lesions, specificity was 55.9% at 98.4% sensitivity. In contrast, BI-RADS category alone yielded a specificity of 27.5% at 97.5% sensitivity, showing the added value of the machine learning classification. Combination of SWE and QUS features yielded a specificity of 16.3% at 97.6% sensitivity.