Enhanced anchor-free pulmonary nodule detection utilizing distribution prediction and task-aligned learning...

Efficient lung cancer screening relies on the automated and precise detection of pulmonary nodules in computed tomography (CT) images. Conventional anchor-based methodologies often rely on pre-defined anchor boxes, resulting in complex design and constrained adaptability to nodules of varying sizes. In contrast, anchor-free models, characterized by their simpler design, have garnered increasing interest. However, existing anchor-free methods may suffer from spatial misalignment across tasks and limited adaptability to diverse nodule shapes.

This study proposes a novel anchor-free model, named DistAlignNet, which integrates distribution prediction and task-aligned learning to address challenges such as spatial misalignment and the diversity of nodule shapes.

Our model utilizes a one-stage object detection architecture, enabling automatic learning of nodule shapes without explicit anchor box definitions. The model is optimized for both classification and localization tasks. In the localization task, we incorporate distribution prediction to improve noise robustness and adapt to the distinctive shape patterns of each nodule. In addition, we introduce a task-aligned learning strategy to mitigate inconsistencies between different tasks and dynamically adjust the label assignment process.

Experimental evaluations were performed on the publicly available LUNA16 dataset, where our model achieved an average Competition performance metric (CPM) score of 0.924. Compared with the SCPM-Net model that reported the strongest performance among the existing approaches, the statistical analysis yielded $t=5.002$ and , with a Cohen's $d$ of 1.582, reflecting a statistically significant improvement. These results show that our model achieves superior performance compared with state-of-the-art anchor-based and anchor-free methods for pulmonary nodule detection.

Quantitative results and qualitative assessments demonstrate the effectiveness of our model, indicating its potential for clinical deployment in pulmonary nodule detection.