Detecting bone lesions in X-ray under diverse acquisition conditions

Area of Science:

• Medical imaging
• Artificial intelligence in radiology
• Oncology

Background:

• Diagnosing primary bone tumors is difficult due to non-specific initial symptoms.
• Early bone cancer detection is vital for patient prognosis.
• Incidental radiographic findings of bone lesions are often overlooked.

Purpose of the Study:

• To develop an automatic algorithm for detecting bone lesions in conventional radiographs for earlier diagnosis.
• To overcome challenges in detecting bone cancer, including low prevalence and data diversity from various imaging sources.

Main Methods:

• Utilized an off-the-shelf object detection algorithm trained on radiographs.
• Implemented a novel preprocessing stage involving self-supervised region-of-interest detection with vision transformers (ViT).
• Applied foreground-based histogram equalization for contrast enhancement in relevant regions.

Main Results:

• Achieved 82.43% sensitivity with a 1.5% false-positive rate.
• Demonstrated superior performance compared to existing preprocessing methods.
• Obtained 82.5% accuracy and an Intersection over Union (IoU) of 0.69 for lesion detection.

Conclusions:

• The proposed preprocessing method effectively handles the diversity of radiographs from health maintenance organizations (HMOs) and emergency departments (EDs).
• Facilitates more reliable automatic analysis of bone lesions in varied clinical settings.