Deep Multi-Objective Learning from Low-Dose CT for Automatic Lung-RADS Report Generation

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Chest radiography interpretation for radiology report generation is time-consuming and prone to errors, especially where radiologists are scarce.
• Lack of expert radiologists and diagnostic expertise in certain regions exacerbates challenges in timely and accurate lung cancer diagnosis.
• Automated systems are needed to support radiologists and improve diagnostic accessibility.

Purpose of the Study:

• To develop and evaluate a multi-objective deep learning model, CT2Rep, for automated generation of lung radiology reports from CT scans.
• To extract key semantic features from lung CT scans for predicting lung cancer indicators.
• To assess the performance and practicality of the CT2Rep model in a clinical context.

Main Methods:

• A multi-objective deep learning model, CT2Rep (Computed Tomography to Report), was proposed.
• The model utilized 458 CT scans, extracting 107 radiomics features and 6 slices of segmentation-related nodule features.
• CT2Rep was designed to simultaneously predict lung nodule position, margin, and texture.

Main Results:

• CT2Rep achieved a remarkable F1-score of 87.29% in predicting key lung cancer indicators.
• A satisfaction survey indicated that 95% of generated reports were rated as satisfactory by medical personnel.
• The model demonstrated robust performance in generating quantitative lung diagnosis reports.

Conclusions:

• CT2Rep shows significant potential for producing reliable quantitative lung diagnosis reports, addressing radiologist shortages and expertise gaps.
• The model can provide crucial diagnostic indicators from lung CT scans, facilitating widespread application in medical settings.
• Automated report generation using CT2Rep can enhance diagnostic efficiency and potentially improve patient outcomes.