Weakly-Supervised Detection of Bone Lesions in CT.

Area of Science:

• Medical Imaging
• Oncology
• Radiology

Background:

• Skeletal metastases are common in breast and prostate cancer.
• Computed Tomography (CT) is used for lesion size measurement, but detection is challenging due to lesion variability.
• Accurate localization of bone lesions is essential for monitoring treatment response.

Purpose of the Study:

• To develop an automated pipeline for detecting bone lesions (lytic, blastic, mixed) in CT volumes.
• To address the need for reliable tracking of metastatic bone disease progression.

Main Methods:

• A proxy segmentation task was employed using weak 3D segmentation masks derived from radiologist-marked 2D slices.
• A 3D full-resolution nnUNet model was trained on these weak annotations for lesion segmentation and detection.
• The pipeline was developed using incomplete and partial training data.

Main Results:

• The automated method achieved a precision of 96.7% in detecting bone lesions on CT scans.
• A recall of 47.3% was obtained, demonstrating the system's ability to identify lesions.
• This pilot work represents the first attempt at direct bone lesion detection in CT via a proxy segmentation approach.

Conclusions:

• The developed automated technique shows promise for detecting bone lesions in CT scans.
• This method can facilitate more reliable tracking of metastatic bone disease.
• Further development with more comprehensive data could enhance detection performance.