Soft label-guided transformer for radiology report generation.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Radiology reports are crucial for physician decision-making.
• Automatic report generation can enhance efficiency but often lacks detail.
• Current methods struggle to accurately describe specific lesion features.

Purpose of the Study:

• To develop an automated radiology report generation method that mimics the radiologist's diagnostic process.
• To improve the accuracy and detail of generated radiology reports, especially for abnormal findings.

Main Methods:

• Proposed a soft label-guided transformer (SLGT) model.
• Employed a soft label-guided attention mechanism to focus on disease-related features.
• Aligned image and text features and used generated text to guide representations.
• Utilized a hybrid loss function for text generation, disease classification, and visual-textual alignment.

Main Results:

• The SLGT model was evaluated on IU X-ray, MIMIC-CXR, and COV-CTR datasets.
• SLGT demonstrated superior performance compared to existing state-of-the-art models across all tested datasets.

Conclusions:

• The developed SLGT model significantly enhances automatic radiology report generation.
• This advancement makes automated reports more feasible for computer-aided diagnosis applications.