Weakly-supervised learning-based pathology detection and localization in 3D chest CT scans.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Advancements in anomaly detection enable new radiological reading assistance tools.
• Clinical adoption requires high sensitivity and low false positive rates.

Purpose of the Study:

• Develop a novel method for identifying multiple chest pathologies using contrastive self-supervised learning.
• Provide 3D localization of abnormalities including low lung density area (LLDA), consolidation (CONS), nodules (NOD), and interstitial pattern (IP).

Main Methods:

• A 3D patch-based classifier with a Resnet backbone encoder, pretrained using SimCLR, was developed.
• The model was fine-tuned on a labeled dataset for classification and localization of four chest abnormalities and normal cases.
• Inference involved generating probability maps for multi-label patient-level prediction and evaluating different training strategies.

Main Results:

• The method achieved an AUROC of 0.931 for multi-label and 0.963 for binary classification.
• High AUROC values were noted for interstitial pattern (0.974) and low lung density area (0.952).
• Contrastive pretraining outperformed ImageNet pretraining and random initialization, with localization completeness at 88.8% and accuracy at 88.3%.

Conclusions:

• The proposed method effectively integrates self-supervised learning for pretraining and a patch-based approach for 3D pathology localization.
• It demonstrates potential for efficient detection and localization of multiple anomalies within a single CT scan.
• The aggregation method enables multi-label prediction at the patient level.