EDICNet: An end-to-end detection and interpretable malignancy classification network for pulmonary nodules in...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Pulmonary nodules are common findings on computed tomography (CT) scans.
• Accurate detection and characterization of these nodules are crucial for early diagnosis and treatment of lung diseases.
• Current diagnostic processes can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To develop an interpretable, end-to-end deep learning tool for pulmonary nodule detection and diagnosis on CT scans.
• To automatically identify nodules and classify their malignancy.
• To generate radiologist-interpretable features for enhanced diagnostic insights.

Main Methods:

• Utilized RetinaNet for nodule detection, trained on the LIDC-IDRI dataset.
• Externally validated the nodule detector on a UCLA dataset.
• Employed a hierarchical semantic convolutional neural network (HSCNN) for nodule malignancy classification and feature extraction.
• Trained and validated the HSCNN using CT data from the UCLA dataset with 5-fold cross-validation and data augmentation.

Main Results:

• Nodule detector achieved an average precision of 0.24 at an IoU of 0.5 on the validation set.
• The HSCNN achieved a mean AUC of 0.89 and a mean accuracy of 0.74 for predicting nodule malignancy.
• Mean accuracies for predicting nodule diameter, consistency, and margin were 0.59, 0.74, and 0.75, respectively.
• The pipeline successfully detects nodules ≥ 5 mm and assigns radiologist-interpretable features.

Conclusions:

• An interpretable, end-to-end deep learning pipeline for pulmonary nodule detection and diagnosis has been developed.
• The system demonstrates promising performance in identifying nodules and classifying malignancy.
• The generated radiologist-interpretable features can potentially improve diagnostic efficiency and consistency in clinical practice.