MTMR-Net

MTMR-Net implements a multi-task deep learning model that performs lung nodule classification and attribute score regression to support radiological diagnosis of lung cancer.

Key Features:

• Multi-Task Learning Framework: Integrates simultaneous nodule classification and attribute score regression to exploit shared information between tasks.
• Margin Ranking Loss: Uses a Siamese network architecture with margin ranking loss to improve discrimination in ambiguous nodule cases.
• Feature Relationship Exploration: Employs recursive feature elimination to iteratively identify and rank features most related to malignancy.
• Performance Validation: Evaluated on the LIDC-IDRI dataset and reported competitive classification performance with more precise attribute scoring compared to existing state-of-the-art methods.

Scientific Applications:

• Automated Radiological Assessment: Provides simultaneous classification and attribute regression outputs for CT lung nodule analysis in radiology contexts.
• Ambiguity Resolution in Diagnosis: Enhances discrimination of ambiguous benign versus malignant nodules to support early lung cancer diagnosis.

Methodology:

The approach employs a multi-task deep learning framework that explicitly models the interdependence between nodule classification and attribute regression; a Siamese network trained with margin ranking loss is used to refine discrimination in challenging cases; recursive feature elimination is applied to identify and rank malignancy-related features, and evaluation was performed on the LIDC-IDRI dataset.