EMeRALDS: Electronic Medical Record Driven Automated Lung Nodule Detection and Classification in Thoracic CT Images

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Oncology

Background:

• Lung cancer is a major cause of mortality globally, often due to late diagnosis.
• Early detection of pulmonary nodules in CT scans is crucial for effective treatment.

Purpose of the Study:

• To develop an advanced computer-aided diagnosis (CAD) system for lung nodule detection and classification.
• To leverage large vision-language models (VLMs) integrated with radiomics and synthetic electronic medical records (EMRs) for enhanced accuracy.

Main Methods:

• An end-to-end CAD pipeline was developed with a detection module (CADe) using Segment Anything Model 2 (SAM2) and a diagnosis module (CADx).
• The CADe module utilized text prompts encoded by CLIP, replacing standard visual prompts for nodule segmentation.
• The CADx module calculated similarity scores between segmented nodules and radiomic features, incorporating clinical context from synthetic EMRs.

Main Results:

• The CADe module achieved a Dice score of 0.92 and IoU of 0.85 for nodule segmentation.
• The CADx module demonstrated 0.97 specificity in malignancy classification, outperforming existing fully supervised methods.
• The system showed strong performance in zero-shot settings for lung nodule analysis on the LIDC-IDRI dataset.

Conclusions:

• The integration of VLMs, radiomics, and synthetic EMRs provides accurate and clinically relevant CAD for pulmonary nodules.
• The proposed system has significant potential to improve early lung cancer detection, diagnostic confidence, and patient management in clinical settings.