HIFUNet: Multi-class Segmentation of Uterine Regions from ...
Fully-automatic AI segmentation of subcortical regions: comparison … · 2020. 3. 4. ·...
1
Fully-automatic AI segmentation of subcortical regions: comparison of ATLAS and deep-learning based approaches Kinnunen K 1 , Weatheritt J 1 , Joules R 1 & Wolz R 1,2 1 IXICO plc, London UK | 2 Department of Computing, Imperial College, London, UK Both methodologies perform well for both regions, in terms of volume and dice metrics. The CNN outperforms LEAP for the caudate, but for the putamen, both are statistically equivalent in terms of performance. Changes in volume of the caudate and putamen are used as biomarkers to track the development of Huntington’s disease and monitor the potential effect of interventional treatments. Accurate volume calculations, obtained via segmentations, are of utmost clinical importance. We compare two AI methodologies for anatomical segmentation, tuned for subcortical regions. A neural network is a set of operations, loosely modelled on the human brain, that by repeated exposure to 100s of examples of labelled data (caudate or putamen), learns how to predict anatomical regions at the pixel level. Neural Network Method LEAP Switch to inference mode Extract 3D box around subcortical region Train 3D-UNET 2 by gradient descent LEAP is an atlas based approach that selects nearest neighbours in manifold space. A majority voting on these atlas's labels fuses into one prediction. Graph cut segmentation is performed on the output to refine the final prediction. Label Transformation Brain Extraction N4 Bias Correction Compute distance values in manifold space Select atlas Label fusion Graph cuts refinement Final prediction Results To test the models on the caudate, we used 100 manually labelled brains from an early manifest HD population. For the putamen, we used semi- automatic labels from 180 brains from ADNI 1 . Putamen Caudate ROI To test the models' efficacy, statistics are computed across the validation set. For the caudate dataset, the two methods perform: 2. Ronneberger, Olaf, et al. "U-net: Convolutional networks for biomedical image segmentation." International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015. 1. Jack Jr, Clifford R., et al. "The Alzheimer's disease neuroimaging initiative (ADNI): MRI methods." Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 27.4 (2008): 685-691. 3. Wolz, Robin, et al. "LEAP: learning embeddings for atlas propagation." NeuroImage 49.2 (2010): 1316-1325. Make predictions For both the caudate and putamen, the neural network outperforms LEAP. Despite the hard boundary with the ventricles, the caudate is a more challenging region than the putamen for LEAP – where as the neural network is comparable for both regions. TRAIN & TEST PIPELINE LEAP CNN Manual Example caudate prediction: LEAP v CNN against manual label ground truth. L-R: moving axially through brain *multiple brains can be run simultaneously at no extra cost, time excludes any pre-registration required encoder decoder trained model DICE Execution Time LEAP 0.8988 (0.0281) 15m 24s CNN 0.9017 (0.00983) 1m* Putamen statistics ixico.com | IXICO | @ IXICOnews | [email protected]
Transcript of Fully-automatic AI segmentation of subcortical regions: comparison … · 2020. 3. 4. ·...
Fully-automatic AI segmentation of subcortical regions: comparison of ATLAS and deep-learning based approaches
Kinnunen K1, Weatheritt J1, Joules R1 & Wolz R1,2
1IXICO plc, London UK | 2Department of Computing, Imperial College, London, UK
Both methodologies perform well for both regions,in terms of volume and dice metrics.
The CNN outperforms LEAP for the caudate, butfor the putamen, both are statistically equivalentin terms of performance.
Changes in volume of the caudate and putamenare used as biomarkers to track the developmentof Huntington’s disease and monitor the potentialeffect of interventional treatments. Accuratevolume calculations, obtained via segmentations,are of utmost clinical importance.We compare two AI methodologies for anatomicalsegmentation, tuned for subcortical regions.
A neural network is a set of operations, loosely modelled on the human brain, that by repeated exposure to 100s of examples of labelled data (caudate or putamen), learns how to predict anatomical regions at the pixel level.
Neural Network Method
LEAP
Switch to inference modeExtract 3D box around subcortical region
Train 3D-UNET2
by gradient descent
LEAP is an atlas based approach that selects nearest neighbours in manifold space. A majority voting on these atlas's labels fuses into one prediction. Graph cut segmentation is performed on the output to refine the final prediction.
Label
Transformation
Brain Extraction N4 Bias Correction
Compute distance values in manifold space
Select atlas Label fusionGraph cutsrefinement
Final prediction
Results
To test the models on thecaudate, we used 100manually labelled brainsfrom an early manifestHD population. For theputamen, we used semi-automatic labels from180 brains from ADNI1.
Putamen
Caudate ROI
To test the models' efficacy, statistics are computed acrossthe validation set. For the caudate dataset, the twomethods perform:
2. Ronneberger, Olaf, et al. "U-net: Convolutional networks for biomedical image segmentation." International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015.
1. Jack Jr, Clifford R., et al. "The Alzheimer's disease neuroimaging initiative (ADNI): MRI methods." Journal of Magnetic Resonance Imaging: An Official Journal of the International Society for Magnetic Resonance in Medicine 27.4 (2008): 685-691.
3. Wolz, Robin, et al. "LEAP: learning embeddings for atlas propagation." NeuroImage 49.2 (2010): 1316-1325.
Make predictions
For both the caudate and putamen, the neural network outperforms LEAP. Despite the hard boundary with the ventricles, the caudate is a more challenging region than the putamen for LEAP – where as the neural network is comparable for both regions.
TR
AIN
& T
ES
T P
IPE
LIN
E
LEAPCNNManual
Example caudate prediction: LEAP v CNN against manual label ground truth. L-R: moving axially through brain
*multiple brains can be run simultaneously at no extra cost, time excludes any pre-registration required
encoder decoder
trained model
DICE Execution Time
LEAP 0.8988(0.0281)
15m 24s
CNN 0.9017(0.00983)
1m*
Putamen statistics
ixico.com | IXICO | @ IXICOnews | [email protected]
