Bone Metastasis Detection at CT with Deep Learning Models Trained Using Multicenter, Multimodal Reference Standards:...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Bone metastases detection on CT scans is crucial for cancer staging and treatment.
• Current detection methods rely on human interpretation, which can be variable and time-consuming.
• Deep learning offers potential for automated and accurate metastasis detection.

Purpose of the Study:

• To develop and validate deep learning models for detecting bone metastases on abdominal and thoracic CT scans.
• To assess the impact of lesion visibility (visible, indeterminate, invisible) on model performance.
• To compare the performance of deep learning models against human readers (musculoskeletal radiologists and radiologists in training).

Main Methods:

• Retrospective multicenter study using CT scans from patients with bone metastases.
• MRI and PET-CT were used as reference standards to categorize lesions based on CT visibility.
• Two nnU-Net deep learning models were trained: Model 1 (CT-visible metastases) and Model 2 (visible and indeterminate metastases).
• Performance evaluated using lesion-level precision/recall and scan-level AUC, compared against human readers.

Main Results:

• Model 2 demonstrated higher recall (41.8%) compared to Model 1 (33.9%) for all lesions.
• Both models showed superior precision compared to radiologists in training (66.6%) and musculoskeletal radiologists (66.5%).
• Model 2 achieved comparable recall (43.8%) and scan-level AUC (0.80) to expert radiologists.

Conclusions:

• A deep learning model trained with multimodal reference standards achieved expert-level performance in detecting bone metastases on body CT.
• The model's ability to incorporate indeterminate lesions enhances its detection capabilities.
• Deep learning holds significant promise for improving the accuracy and efficiency of bone metastasis detection in clinical practice.