A supervised case-based reasoning approach for explainable thyroid nodule diagnosis

As an explainable experience-based artificial intelligence technique, case-based reasoning (CBR) has been widely used to help diagnose many diseases, but the application of CBR in the diagnosis of thyroid nodules (TDNs) is rarely studied. To fill this research gap, this paper proposes a supervised CBR approach to help diagnose TDNs. The proposed approach first investigates the correlation between the feature diagnoses of historical TDN cases and the corresponding overall diagnoses using the canonical correlation analysis technique. Then the learned canonical variables are used to reconstruct TDN cases. Based on the reconstructed historical case base, a classifier is constructed to provide pathological diagnosis predictions for new TDN cases. To explain these predictions with similar historical TDN cases, a convex optimization model is constructed to determine the similarity between historical TDN cases and new TDN cases. Finally, a weighted combination scheme is designed to generate an explainable pathological diagnosis for each new TDN case based on its similar historical TDN cases. The proposed approach not only avoids the burdensome parameter tuning task but also reduces the likelihood of retrieving noisy historical cases as similar cases of new cases with a supervised case retrieval process. Using a real diagnostic dataset collected from the ultrasound department of a local hospital, the effectiveness of the proposed approach in diagnosing TDNs is validated and its advantages are further highlighted by comparison with the traditional CBR approach and six mainstream machine learning models. • A supervised case-based approach (CBR) is proposed for explainable diagnosis of thyroid nodules. • Both case features and case solutions are considered to determine the similarity between different cases. • Predictions of unexplainable machine learning models are explained using similar historical cases. • The proposed approach is compared with traditional CBR approach and mainstream machine learning models.