Evaluation of a novel deep learning-based classifier for perifissural nodules

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Pathology

Background:

• Perifissural nodules (PFNs) are common findings on chest CT scans.
• Accurate classification of PFNs is crucial for lung cancer screening and diagnosis.
• Current classification methods rely on expert radiologist interpretation, which can be time-consuming and subject to inter-observer variability.

Purpose of the Study:

• To evaluate the diagnostic performance of a novel convolutional neural network (CNN) for classifying typical perifissural nodules (PFNs).
• To compare the CNN's classification accuracy against the consensus of experienced radiologists.

Main Methods:

• A dataset of 1668 unique pulmonary nodules from two international centers was utilized.
• Nodules were classified as typical PFN, atypical PFN, or non-PFN by three expert readers.
• A PFN-specific CNN was trained and validated on a subset of the data, with performance assessed on a held-out test set using ROC curves, confusion matrices, and Cohen's kappa.

Main Results:

• The PFN-CNN achieved an Area Under the ROC Curve (AUC) of 95.8% on the test set.
• Sensitivity was 95.6% and specificity was 88.1% for identifying typical PFNs.
• The agreement between the CNN and the consensus of readers (k = 0.62-0.75) was comparable to the inter-reader agreement (k = 0.64-0.79).

Conclusions:

• The developed PFN-CNN demonstrates excellent performance in classifying typical PFNs.
• The CNN's classification agreement with expert radiologists falls within the expected range of inter-reader variability.
• This AI-driven system shows significant potential for clinical application in lung nodule screening, assisting in the efficient exclusion of perifissural nodules.