Related Experiment Video
Updated: Jun 23, 2025
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
An explainable machine learning model to solid adnexal masses diagnosis based on clinical data and qualitative
Annarita Fanizzi1, Francesca Arezzo2,3, Gennaro Cormio2,4
1Laboratorio Biostatistica e Bioinformatica, I.R.C.C.S. Istituto Tumori 'Giovanni Paolo II', Bari, Italy.
Background:
Accurate characterization of newly diagnosed a solid adnexal lesion is a key step in defining the most appropriate therapeutic approach. Despite guidance from the International Ovarian Tumor Analyzes Panel, the evaluation of these lesions can be challenging. Recent studies have demonstrated how machine learning techniques can be applied to clinical data to solve this diagnostic problem. However, ML models can often consider as black-boxes due to the difficulty of understanding the decision-making process used by the algorithm to obtain a specific result.
Aims:
For this purpose, we propose an Explainable Artificial Intelligence model trained on clinical characteristics and qualitative ultrasound indicators to predict solid adnexal masses diagnosis.
Materials & Methods:
Since the diagnostic task was a three-class problem (benign tumor, invasive cancer, or ovarian metastasis), we proposed a waterfall classification model: a first model was trained and validated to discriminate benign versus malignant, a second model was trained to distinguish nonmetastatic versus metastatic malignant lesion which occurs when a patient is predicted to be malignant by the first model. Firstly, a stepwise feature selection procedure was implemented. The classification performances were validated on Leave One Out scheme.
Results:
The accuracy of the three-class model reaches an overall accuracy of 86.36%, and the precision per-class of the benign, nonmetastatic malignant, and metastatic malignant classes were 86.96%, 87.27%, and 77.78%, respectively.
Discussion:
SHapley Additive exPlanations were performed to visually show how the machine learning model made a specific decision. For each patient, the SHAP values expressed how each characteristic contributed to the classification result. Such information represents an added value for the clinical usability of a diagnostic system.
Conclusions:
This is the first work that attempts to design an explainable machine-learning tool for the histological diagnosis of solid masses of the ovary.
