Exploiting Multi-Head Attention Maps Into A Deep Riemannian Representation to Quantify Pulmonary Nodules

Nodules are the principal biomarker of lung cancer, observed in CT scans as masses with abnormal variations in size and texture. Today, the radiological finding characterization and disease stratification remain highly subjective and expert-dependent. Deep representations have recently addressed diagnosis support, but their effectiveness strongly depends on a significant amount of balanced and annotated data. This work introduces a Riemannian mechanism that fully exploits attention map outputs, dealing with textural nodule variability in scarce data scenarios. The end-to-end strategy first recovers non-local textural relationships from multi-head attention modules. Then, using noise-robust statistics, a symmetric positive embedding (SPD) captures pairwise relationships between nodule patterns. The resultant SPD embeddings are mapped to deep geometrical modules to learn compact and discriminative descriptors with respect to the disease. The proposed approach was validated on the LIDC-IDRI dataset demonstrating competitive results in the SOA with a 91.83% on F1-score, and 95.19% on AUC.