Attribute and Malignancy Analysis of Lung Nodule on Chest CT with Cause-and-Effect Logic

Lung cancer is the leading cause of cancer-related death. Early detection and treatment are crucial to improve survival rates. Radiologists determine whether the nodules are benign or malignant by observing their morphological attributes. However, this can be a challenging task for well-trained doctors.

We propose a more efficient automatic lung nodule analysis method, which establishes a clear cause-and-effect logic relationship between attribute features and malignancy features by incorporating multiple instance learning (MIL). The designed MIL classifier aggregates the learned instance weights and corresponding attribute features to form malignancy features. Compared to existing methods, it starts by mirroring the way radiologists observe nodules, then proceeds to extract the multi-scale morphological attribute characteristics of the nodules. The instance weight also serves as the attribute score of the attribute, providing a reference for consultation.

Our method was validated using the LIDC-IDRI dataset and achieved an accuracy of 93.05% on benign-malignant classification task with the added capability of accurately scoring the attributes.

The proposed method based on attribute score regression and multi-instance learning establishes the causal relationship between attribute scores and malignancy. This method improves accuracy in nodule classification and addresses the issue of poor model interpretability.