Robust explanation supervision for false positive reduction in pulmonary nodule detection

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Radiology

Background:

• Lung cancer is a leading cause of cancer death, often diagnosed late due to subtle early symptoms.
• Pulmonary nodules detected via thoracic CT scans are crucial indicators for early lung cancer diagnosis and improved survival rates.
• Current radiologist-based analysis of CT images for nodules is prone to errors, necessitating advanced diagnostic tools.

Purpose of the Study:

• To develop an explainable AI (XAI) framework for accurate pulmonary nodule detection in CT images.
• To enhance the understanding between deep learning (DL) algorithms and radiologists in identifying cancerous nodules.
• To integrate XAI methods to improve the reliability and interpretability of DL-based nodule detection.

Main Methods:

• Proposed a robust and explainable detection (RXD) framework utilizing explanation supervision with radiologist-annotated nodule contours.
• Implemented imputation methods to reduce noise in human annotations and ensure robust model attributions.
• Trained and validated the framework on thoracic CT image sets from the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset.

Main Results:

• The RXD framework demonstrated consistent improvements in classification performance (AUC) and explanation quality (IoU) with increasing training samples.
• A learnable imputation kernel improved Intersection over Union (IoU) by 24.0% to 80.0%, while a Gaussian imputation kernel achieved an 118.8% improvement over baseline.
• The proposed method showed less performance degradation on smaller datasets and better alignment with expert opinions compared to baseline models.

Conclusions:

• A novel pulmonary nodule detection framework integrating robust explanation supervision (RES) was successfully demonstrated.
• The framework enhances nodule classification and morphology assessment, aiding early lung cancer diagnosis.
• This AI approach can potentially reduce radiologist workload, allowing greater focus on diagnosing and prognosing potentially cancerous pulmonary nodules.