Pathomics-based machine learning models for optimizing LungPro navigational bronchoscopy in peripheral lung lesion...

Area of Science:

• Pulmonary Medicine
• Artificial Intelligence in Medicine
• Digital Pathology

Background:

• LungPro navigational bronchoscopy is used for diagnosing peripheral pulmonary lesions.
• Accurate diagnosis is crucial for effective management, especially for lesions initially appearing negative.
• Pathomics, analyzing image features from histopathology, offers potential for improved diagnostic accuracy.

Purpose of the Study:

• To develop a pathomics-based machine learning model to enhance LungPro diagnostic efficacy.
• To optimize management strategies for peripheral pulmonary lesions with negative LungPro results.
• To integrate clinical, imaging, and pathomic data for a multimodal diagnostic framework.

Main Methods:

• Collected clinical data and H&E-stained WSIs from 144 patients undergoing LungPro bronchoscopy.
• Developed and validated an AI model using CNN and MIL for feature extraction and aggregation.
• Integrated clinical, imaging, and pathomic data into a multimodal framework evaluated on 50 LungPro-negative patients.

Main Results:

• Age, lesion boundary, and CT attenuation were independent risk factors for malignant lesions.
• The MIL fusion model achieved an AUC of 0.792 (training) and 0.777 (test) for lung cancer diagnosis.
• The multimodal framework improved diagnostic yield to 0.848 and correctly identified 71.43% of malignant lesions initially negative by LungPro.

Conclusions:

• A fusion diagnostic model incorporating clinical and pathomic features significantly enhances LungPro diagnostic accuracy.
• The model aids in detecting subtle malignant characteristics in peripheral pulmonary lesions.
• This approach supports precise therapeutic interventions for lesions initially classified as negative by LungPro.