A human-interpretable machine learning pipeline based on ultrasound to support leiomyosarcoma diagnosis

• Published in: Artificial intelligence in medicine Vol. 146; p. 102697
• Main Authors: Lombardi, Angela, Arezzo, Francesca, Di Sciascio, Eugenio, Ardito, Carmelo, Mongelli, Michele, Di Lillo, Nicola, Fascilla, Fabiana Divina, Silvestris, Erica, Kardhashi, Anila, Putino, Carmela, Cazzolla, Ambrogio, Loizzi, Vera, Cazzato, Gerardo, Cormio, Gennaro, Di Noia, Tommaso
• Format: Journal Article
• Language: English
• Published: Netherlands 01.12.2023
• Subjects: Algorithms; Humans; Leiomyosarcoma - diagnostic imaging; Machine Learning; Ultrasonography; Interpretability; CAD; Human-centered AI; Leiomyosarcoma; Machine learning; eXplainable artificial intelligence; Ultrasound
• ISSN: 0933-3657; 1873-2860
• DOI: 10.1016/j.artmed.2023.102697

The preoperative evaluation of myometrial tumors is essential to avoid delayed treatment and to establish the appropriate surgical approach. Specifically, the differential diagnosis of leiomyosarcoma (LMS) is particularly challenging due to the overlapping of clinical, laboratory and ultrasound features between fibroids and LMS. In this work, we present a human-interpretable machine learning (ML) pipeline to support the preoperative differential diagnosis of LMS from leiomyomas, based on both clinical data and gynecological ultrasound assessment of 68 patients (8 with LMS diagnosis). The pipeline provides the following novel contributions: (i) end-users have been involved both in the definition of the ML tasks and in the evaluation of the overall approach; (ii) clinical specialists get a full understanding of both the decision-making mechanisms of the ML algorithms and the impact of the features on each automatic decision. Moreover, the proposed pipeline addresses some of the problems concerning both the imbalance of the two classes by analyzing and selecting the best combination of the synthetic oversampling strategy of the minority class and the classification algorithm among different choices, and the explainability of the features at global and local levels. The results show very high performance of the best strategy (AUC = 0.99, F1 = 0.87) and the strong and stable impact of two ultrasound-based features (i.e., tumor borders and consistency of the lesions). Furthermore, the SHAP algorithm was exploited to quantify the impact of the features at the local level and a specific module was developed to provide a template-based natural language (NL) translation of the explanations for enhancing their interpretability and fostering the use of ML in the clinical setting.