A multi-attention deep architecture to stratify lung nodule malignancy from CT scans

Abstract

Lung cancer remains the principal cause of cancer-related deaths. Nodules are the main radiological finding, typically observed from low-dose CT scans. Nonetheless, the nodule characterization diagnosis remains subjective, reporting a moderate agreement among experts' observations, especially in identifying malignancy stratification. The proposed approach presents a deep multi-attention strategy, validated exhaustively to classify nodule masses according to four malignancy degrees. This work introduces a multi-attention architecture dedicated to stratifying nodules among malignancy stages. The architecture receives volumetric nodule regions and learns multi-scale saliency maps, focusing on determinant malignancy patterns of the observed masses. Specialized attention heads capture related patterns associated with lobulated, textural, and spiculated features. Validation includes an extensive analysis regarding multiple attention features, allowing to establish a correlation with other radiological findings. The proposed approach achieves an AUC of 85.35% for a classical multi-classification and a mean AUC of 82.90% in a one-vs-all validation methodology, showing competitive results in the state-of-the-art. The introduced architecture has capabilities to support nodule stratification and to classify nodule features. The exhaustive validation also suggests a proper generalization performance, which is a potential property to transfer this strategy in real scenarios.