An optimized radiomic approach to breast ultrasound for lesion localization and malignancy classification

Background: Breast cancer remains a major health issue for women worldwide; is it necessary to review the traditional medical workflow to reduce its mortality effect? For sure, Radiomics is fundamental for adding information to medical diagnosis, and the use of Artificial Intelligence (AI) models can conduct a more in-depth analysis of a patient’s health. Objective: This paper aims to optimize the Radiomics workflow for breast ultrasound images by developing a systematic approach that localizes breast cancer lesions and directly assesses their severity from images. The approach involves constructing a square grid to extract radiomic features and simultaneously optimize (i) breast cancer localization, (ii) classification, and (iii) investigate which image features are most strongly correlated with lesion positioning. Methods: Ultrasound images are preprocessed with the application of a square grid k × k . Each grid cell contributes to the extraction of radiomic features, which are then used to locate the lesion and identify its BI-RADS grade. Results: The method localizes with an Intersection Over Union of 78% with k = 4 and a Coverage of 91% with k = 3 . Classification of the severity produces an AUC of 80% thanks to the FIRST order radiomics class. Conclusion: Sometimes, to benefit the explainability of the automatic decision made by AI algorithms, it may be favorable not to rely completely on numerical metrics but also to evaluate their overall quality. For this reason, we developed a new metric, called Coverage, which measures the lesion pattern identified in the image.