Boosting thyroid nodule diagnosis through ultrasound and molecular imaging integration with unsupervised learning.
Authors
Affiliations (5)
Affiliations (5)
- Bicocca Bioinformatics Biostatistics and Bioimaging B4 Center, Department of Medicine and Surgery, University of Milano-Bicocca, Milano, Italy.
- Department of Medicine and Surgery, Pathology Unit, IRCCS Fondazione San Gerardo dei Tintori, University of Milano-Bicocca, Monza, Italy.
- Proteomics and Metabolomics Unit, Department of Medicine and Surgery, University of Milano-Bicocca, Monza, Italy.
- Department of Informatics, Systems and Communication, University of Milano-Bicocca, Milano, Italy.
- Biostatistics and Clinical Epidemiology, Fondazione IRCCS San Gerardo dei Tintori, Monza, Italy.
Abstract
In recent years, matrix-assisted laser desorption-ionization (MALDI) mass spectrometry imaging (MSI) has been applied to cytological thyroid specimens as a complementary tool for the diagnosis of thyroid nodules. Specifically, MALDI-MSI has been proven to be effective for classifying "indeterminate for malignancy" reports. In this work, we analyse the effectiveness of unsupervised methods to further explore whether unsupervised learning (UL) can reveal latent structures within MALDI-MSI, clinical, and ultrasound (US) data that may help better understand the heterogeneity of thyroid nodules, integrating different clinical and molecular information, such as demographic and echographic data. This retrospective study involved 222 patients who underwent US-guided fine-needle aspiration (FNA), with MALDI-MSI analysis. This first dataset (named "Dataset MSI") comprises demographic, molecular, and clinical information. A second dataset of 82 patients ("Dataset MSI + ECHO") was extracted from the first one, containing additional information regarding the US characterization of the thyroid nodules. Unsupervised clustering was performed for each dataset through three distinct methods: k-means, partitioning around medoids (PAM), and hierarchical clustering (HC), and compared to MALDI-MSI classifications. Our results highlight the potential value of unsupervised approaches in exploring the underlying structure of the data. In Dataset MSI, the clustering analysis revealed patterns partially consistent with clinical outcomes and helped group cases that were inconclusive in MALDI-MSI reports, with 91% (20 out of 22) of patients without a clear MALDI-MSI result falling into clinically coherent clusters. Similarly, in Dataset MSI + ECHO, we observed a drastic increase in the overall sensitivity from 0.4 to 0.95 compared to the MALDI-MSI prediction. We also correctly clustered 94% of TIR4/Thy4 (5 out of 5) and TIR5/Thy5 (11 out of 12) patients that MALDI-MSI misclassified. This research highlights the need to incorporate clinical, US, and molecular information into a single learning method in routine diagnosis, especially when US variables are available. This approach may represent a useful exploratory framework to investigate the biological and clinical heterogeneity of thyroid nodules and to guide future studies, potentially leading to more accurate and timely diagnoses. These findings pave the way for further research and applications in diagnosing thyroid nodules.