GOPI-Segm: Deep Learning for Prostate Segmentation
Study Details
Study Description
Brief Summary
Because the diagnostic criteria for prostate cancer are different in the peripheral and the transition zone, prostate segmentation is needed for any computer-aided diagnosis system aimed at characterizing prostate lesions on magnetic resonance (MR) images. Manual segmentation is time consuming and may differ between radiologists with different expertise. We developed and trained a convolutional neural network algorithm for segmenting the whole prostate, the transition zone and the anterior fibromuscular stroma on T2-weighted images of 787 MRIs from an existing prospective radiological pathological correlation database containing prostate MRI of patients treated by prostatectomy between 2008 and 2014 (CLARA-P database).
The purpose of this study is to validate this algorithm on an independent cohort of patients.
Condition or Disease | Intervention/Treatment | Phase |
---|---|---|
|
Study Design
Arms and Interventions
Arm | Intervention/Treatment |
---|---|
Patients with a MRI on a 3 Tesla (T) unit The total validation cohort is composed of axial T2-weighted images of the prostate obtained from 31 prostate MRIs on a 3T unit randomly chosen among the prostate MRIs performed at the Hospices Civils de Lyon in 20162015-2019 |
Other: Comparison of prostate multi-zone segmentation obtained with an automatic deep learning-based algorithm and two expert radiologists
The algorithm is used to perform a multizone segmentation of the prostate including delineation of : the whole prostate contours, the transition zone contours, the anterior fibromuscular stroma.
The contours is independently corrected by 2 radiologists. The corrected contours of the different zones will be stored and for each zone 6 different metrics will be used to evaluate the difference between the initial and corrected contours:
Mean Mesh Distance: Average Boundary Distance (ABD) for each point of the reference segmentation. The distance to the closest point of the compared segmentation is first computed. Then the average of all these distances is computed and gives the ABD
General Hausdorff distance (HD)
95% percentile (P) of the HD and the 95th (P) of the asymmetric HD distribution
95% HD modified (HD95_1): different approach by first computing the 95th (P) of the asymmetric HD then taking the maximum
Dice coefficient
Difference in volumes
|
Patients with a MRI on a 1.5 Tesla unit The total validation cohort is composed of axial T2-weighted images of the prostate obtained from 31 prostate MRIs on a 1.5T unit randomly chosen among the prostate MRIs performed at the Hospices Civils de Lyon in 20162015-2019 |
Other: Comparison of prostate multi-zone segmentation obtained with an automatic deep learning-based algorithm and two expert radiologists
The algorithm is used to perform a multizone segmentation of the prostate including delineation of : the whole prostate contours, the transition zone contours, the anterior fibromuscular stroma.
The contours is independently corrected by 2 radiologists. The corrected contours of the different zones will be stored and for each zone 6 different metrics will be used to evaluate the difference between the initial and corrected contours:
Mean Mesh Distance: Average Boundary Distance (ABD) for each point of the reference segmentation. The distance to the closest point of the compared segmentation is first computed. Then the average of all these distances is computed and gives the ABD
General Hausdorff distance (HD)
95% percentile (P) of the HD and the 95th (P) of the asymmetric HD distribution
95% HD modified (HD95_1): different approach by first computing the 95th (P) of the asymmetric HD then taking the maximum
Dice coefficient
Difference in volumes
|
Outcome Measures
Primary Outcome Measures
- Mean Mesh Distance (Mean) between the contours of the whole prostate made by the algorithm and the two radiologists [Month 11]
The Mean Mesh Distance corresponds to the Average Boundary Distance (ABD) for each point of the reference segmentation. The distance to the closest point of the compared segmentation is first computed. Then the average of all these distances is computed and gives the ABD. The Mean Mesh Distance between the contours of the whole prostate made by the algorithm and each radiologist will be used as primary outcome measure.
Eligibility Criteria
Criteria
Inclusion Criteria:
-
Prostate MRI contained in the PACS of the Hospices Civils de Lyon
-
Performed in 2016-2019
Exclusion Criteria:
- MRIs from patients who already had treatment for prostate cancer
Contacts and Locations
Locations
Site | City | State | Country | Postal Code | |
---|---|---|---|---|---|
1 | Hôpital Edouard Herriot | Lyon | France | 69008 |
Sponsors and Collaborators
- Hospices Civils de Lyon
Investigators
None specified.Study Documents (Full-Text)
None provided.More Information
Publications
None provided.- GOPI-Segmentation_2019