Convolutional neural networks for segmentation of 49 selected bones in CT images show high reproducibility

Artikel i vetenskaplig tidskrift, 2017

Aim: An automated method to calculate Bone Scan Index (BSI) from bone scans has recently been established as a first imaging biomarker in patients with metastatic prostate cancer. BSI has shown to be an independent predictor of survival. PET/CT is more accurate than bone scans in detecting bone metastases.
We therefore decided to develop an automated PET/CT based imaging biomarker for assessment of tumor burden in bone. The aim of this project was to develop a method for automated segmentation and volume calculation of bones in CT images, which is the first step in the process of developing a PET/CT based imaging biomarker.
Materials and Methods: Convolutional neural networks (CNN) were trained to segment 49 selected bones (12 thoracic vertebrae, 5 lumbar vertebrae, sacrum, 2 hip bones, 24 ribs, 2 scapulae, 2 clavicles and the sternum) using manual segmentations in CT images from 23 patients performed by experienced image readers. Anatomical landmarks were detected using a CNN and pruned using a shape model.
These landmarks and the CT image were fed to a second CNN, segmenting the 49 selected bones. After the training process, the CNN segmented the bones in CT images in a separate validation group consisting of 46 patients with prostate cancer. All patients had undergone both 18F-Choline and 18F-NaF PET/CT within a time frame of 3 weeks as part of a previous research project. The two CT scans from each patient were segmented by the CNN and the two volumes of each bone were calculated.
Results: The total volume of the 49 bones was on average 3,086 mL in the 46 patients. The individual bones ranged in volume from 8 mL (left 12th rib) to 440 mL (left hip bone). The reproducibility measured as ratio volume difference/mean volume was on average less than 2% for all bones except for the ribs. The mean volumes, differences and reproducibility for the bones of five anatomical regions were as follow: thoracic vertebrae 39mL, 0.6mL, 1.5%; lumbar vertebra 71mL, 0.8 mL, 1.2%; sacrum, hip bones 386mL, 0.9mL, 0.3%; ribs 26mL, 2.0mL, 8.5%; scapulae, clavicles, sternum 97mL, -0.1mL, -0.4%.
Conclusion: Our CNN based method for automated segmentation of bones in CT images showed high reproducibility. A reproducible way to segment the skeleton and to measure the bone volume will be important in the development of a PET index relating volumes of abnormal PET tracer uptake to the bone volume.