Analytical validation of an automated method for segmentation of the prostate gland in CT images

Artikel i vetenskaplig tidskrift, 2017

Aim: Uptake of PET tracers in the prostate gland may serve as guidance for management of patients with prostate cancer. PET studies alone do, however, not allow for accurate segmentation of the gland, instead the corresponding CT images contain the required anatomical information. Our long-term aim is to develop an objectively measured PET/CT imaging biomarker reflecting PET tracer uptake. In this study we take the first step and develop and validate a completely automated method for 3D-segmentation of the prostate gland in CT images.
Methods: A convolutional neural network (CNN) was trained to segment the prostate gland in CT images using manual segmentations performed by a radiologist in a group of 100 patients, who had undergone 18F-FDG PET/CT. After the training process, the CNN automatically segmented the prostate gland in CT images in a separate validation group consisting of 45 patients with prostate cancer. All patients had undergone a 18F-choline PET/CT as part of a previous research project. The CNN segmentations were compared to manual segmentations performed independently by two radiologists. The volume of the prostate gland was calculated based on segmentations by the CNN and radiologists. The Sørensen-Dice index was used to analyse the overlap between the segmentations by the CNN and the two radiologists.
Results: The prostate volumes were on average 79mL (range 9-212mL) in the 45 patients, measured as mean volumes for the two radiologists. The mean difference in prostate volumes between the two radiologists was 14mL (SD
29mL). The mean volume difference between the CNN segmentation and the mean values from the two radiologists was 22mL (SD 43mL). For the subgroup of patients with prostate volumes <100 mL (n=36), the difference between the radiologists was 9mL (SD 11mL) compared to difference CNN vs radiologists of 7mL (SD 15mL). The Sørensen-Dice index was 0.69 and 0.70 for the comparison between CNN segmentation and the two radiologists, respectively and 0.83 for the comparison between the two radiologists. The corresponding Sørensen-Dice index in the 36 patients with volumes <100 mL were 0.74, 0.75 and 0.83, respectively 
Conclusion: Our CNN based method for automated segmentation of the prostate gland in CT images show good agreement with the corresponding manual segmentations by two radiologists especially for prostade glands with a volume less than 100 mL.