Beyond hopes and hype, the clinical applicability of artificial intelligence decision support tools deserves to be rigorously explored in multicenter settings. With this work, we aimed to do so by focusing our attention on solid breast cancer lesions as detected by ultrasound. Indeed, the journey of a patient usually begins with this first-line imaging modality and fast, non-invasive and cost-effective classification of benign and malignant lesions would be crucial to timely and properly guide subsequent management. For this classification task, we choose a Random Forest ensemble algorithm which showed an accuracy of 82%, not significantly different from that of a dedicated breast radiologist (79.4%, p = 0.815). Interestingly, the radiologist’s performance improved when provided the predictions of the classifier (80.2%), although statistical significance was not reached (p = 0.508). While encouraging, these results still require further validation. In particular, a wider validation in additional centers should be performed for our model, as well as an integration of this software within the clinical workflow. In turn, these developments would allow a prospective assessment of its impact on patient clinical management and outcome.
- Machine learning showed good accuracy in discriminating benign from malignant breast lesions.
- The machine learning classifier’s performance was comparable to that of a breast radiologist.
- The radiologist’s accuracy improved with machine learning, but not significantly.
Authors: Valeria Romeo, Renato Cuocolo, Roberta Apolito, Arnaldo Stanzione, Antonio Ventimiglia, Annalisa Vitale, Francesco Verde, Antonello Accurso, Michele Amitrano, Luigi Insabato, Annarita Gencarelli, Roberta Buonocore, Maria Rosaria Argenzio, Anna Maria Cascone, Massimo Imbriaco, Simone Maurea & Arturo Brunetti