Zero-Shot Lung Disease Detection Using Radiological Symptomatic Descriptors and Pretrained Neural Networks.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Natural Language Processing

Background:

• Accurate zero-shot disease recognition in chest radiography is hindered by challenges in aligning radiological features with clinical text.
• Existing models struggle with generalization to unseen pathologies.

Purpose of the Study:

• To develop and evaluate a novel framework, DVLM (Dual-Head Vision-Language Model with Neural Memory), for enhanced zero-shot disease recognition in chest radiography.
• To improve the model's ability to generalize to new diseases using a neural memory module.

Main Methods:

• Proposed DVLM framework combining Vision Transformer visual encoding with ClinicalBERT text processing.
• Utilized parallel contrastive and supervised learning branches with a neural memory module for pattern storage.
• Evaluated on CheXpert, MIMIC-CXR, and PadChest datasets using multi-seed validation and ablation studies.

Main Results:

• Achieved 90.0% macro-averaged AUROC on CheXpert, with the neural memory module improving performance by +3.3% (p < 0.001).
• Reached 73.5% AUROC for zero-shot classification (25% held-out diseases), outperforming MedKLIP by 2.3%.
• Demonstrated reduced calibration error (72%), high localization accuracy (IoU 0.642), and equitable performance across demographic groups.

Conclusions:

• DVLM shows strong ranking capabilities for triage applications in medical imaging.
• Threshold-based classification for rare diseases requires further improvement and radiologist confirmation for clinical deployment.