Deep Learning Based Staging of Bone Lesions From Computed Tomography Scans.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer commonly metastasizes to bone, necessitating accurate characterization of bone lesions.
• Distinguishing benign from malignant bone lesions is crucial for effective patient management and treatment planning.
• Computed tomography (CT) scans are widely used for detecting bone metastases, but accurate classification remains challenging.

Purpose of the Study:

• To develop and evaluate an efficient deep learning-based classification strategy for characterizing metastatic bone lesions in prostate cancer patients using CT scans.
• To introduce a comprehensive dataset of annotated bone lesions for training and validation.
• To investigate the impact of various lesion features and deep learning algorithms on classification performance.

Main Methods:

• A dataset of 2,880 annotated bone lesions from 114 prostate cancer patients' CT scans was utilized.
• Patient-level stratification and lesion-aware distribution were employed for training, validation, and testing splits (75%/12%/13%).
• Multiple deep learning models, including 2D and 3D Convolutional Neural Networks (CNNs) and their ensembles, were explored, analyzing lesion texture, morphology, size, location, and volumetric information.

Main Results:

• An ensemble model combining 2D ResNet-50 and 3D ResNet-18 achieved the highest accuracy of 92.2% in classifying benign versus malignant bone lesions.
• Lesion texture, volumetric information, and morphology were identified as the most discriminative features for classification.
• The study achieved state-of-the-art lesion-level accuracy on a comprehensive dataset.

Conclusions:

• The developed deep learning strategy demonstrates high accuracy and reliability in classifying bone lesions in prostate cancer patients.
• The findings suggest significant potential for clinical translation, particularly in the early detection and management of bone metastases.
• This approach offers a promising tool for improving the diagnostic accuracy of bone lesion characterization in clinical practice.