Towards Precision Medicine in Sinonasal Tumors: Low-Dimensional Radiomic Signature Extraction from MRI

Abstract

Background: Sinonasal tumors are rare, accounting for 3-5% of head and neck neoplasms. Machine learning (ML) and radiomics have shown promise in tumor classification, but current models lack detailed morphological and textural characterization. Methods: This study analyzed MRI data from 145 patients (76 malignant and 69 benign) across multiple centers. Radiomic features were extracted from T1-weighted (T1-w) images with contrast and T2-weighted (T2-w) images based on manually annotated tumor volumes. A dedicated ML pipeline assessed the effectiveness of different radiomic features and their integration with clinical variables. The DNetPRO algorithm was used to extract signatures combining radiomic and clinical data. Results: The results showed that ML classification using both data types achieved a median Matthews Correlation Coefficient (MCC) of 0.60 ± 0.07. The best-performing DNetPRO models reached an MCC of 0.73 (T1-w + T2-w) and 0.61 (T1-w only). Key clinical features included symptoms and tumor size, while radiomic features provided additional diagnostic insights, particularly regarding gray-level distribution in T2-w and texture complexity in T1-w images. Conclusions: Despite its potential, ML-based radiomics faces challenges in clinical adoption due to data variability and model diversity. Standardization and interpretability are crucial for reliability. The DNetPRO approach helps explain feature importance and relationships, reinforcing the clinical relevance of integrating radiomic and clinical data for sinonasal tumor classification.

Keywords: feature selection; machine learning; medical image analysis; otorhinolaryngology; radiomic.

Figure 1

Scheme of the proposed ML pipelines. Outline of the proposed pipelines. Starting from the MR images, an expert clinician performs the manual identification and segmentation of the tumor volume on which the radiomic features were extracted. The data were split into train–test (and validation) according to a cross-validation scheme, and the different feature selection approaches (standard machine learning or DNetPRO) were applied to feed the final classification model.

Figure 2

Radiomic embedding analysis. PacMAP projections of the radiomic data extracted from CE T₁-w, T₂-w, and CE T₁-w and T₂-w images. For each combination of features, we colored the data points according to the center of provenance (Varese and Como, first row) and in relation to the cancer type (Malignant and Benign, second row). The clustering of the points was performed using the HDBSCAN algorithm: for the sake of readability, the cluster labels were reported only in the upper row, associated with the kernel density estimations of the cluster boundaries.

Figure 3

Distributions of performances stratified according to the different analyses. Starting from the top, we reported the results obtained by considering the combination of CE T₁-w and T₂-w radiomic features, the CE T₁-w-only subset, and the T₂-w-only subset. In the left column, we reported the performances obtained by the application of a standard ML pipeline on the clinical, radiomics, and clinical and radiomics combinations of data. In the central column of the figure, we reported the performances obtained considering each radiomic group of variables individually. The right column shows the performance obtained by the application of the DNetPRO algorithm, split in relation to the two possible procedures. For each plot, we highlighted with a dashed red line the median score obtained by considering only the clinical values, keeping it as reference for the radiomic contribution in the classification task.

Figure 4

DNetPRO best signatures. Network representations of the best signatures identified by the DNetPRO application, according to procedure A (first row) and procedure B (second row), stratified according to the radiomic groups of features (Wavelet, LoG, and Original). The links between features consider positive/negative synergies of their mutual informative power in the classification task. The node size is represented proportionally to the degree of centrality in the network structure, identifying their importance and ability to cooperate with the other features of the signature network. For the sake of readability, in each plot we highlighted the names of the most central features; the full list of features in each signature and the corresponding network structure are reported in the Supplementary Materials.