MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

Hippocampal Segmentation From CT Scans with a Convolutional Nerual Network

E Porter1*, P Fuentes2 , Z Siddiqui3 , A Thompson3 , T Guerrero3 , (1) Wayne State University, Detroit, MI, (2) Oakland University William Beaumont School of Medicine, Rochester, MI, (3) Beaumont Health, Royal Oak, MI

Presentations

(Wednesday, 7/17/2019) 10:00 AM - 10:30 AM

Room: Exhibit Hall | Forum 2

Purpose: Preliminary results from NRG-CC001 phase III have shown decreased cognitive function failure when hippocampal avoidance is used during whole brain radiotherapy (HA-WBRT). The phase II study, RTOG 0933, found 24% of pre-enrollment contours were unacceptable. Furthermore, conventional human contouring requires MR images, which are unavailable for certain patients. By creating a CT trained hippocampus segmentation model which rivals inter-observer variance, HA-WBRT can be available to all patients.

Methods: A retrospective cohort of patients treated for trigeminal neuralgia was chosen because of the high-resolution CT and MR imaging and the improved accuracy of rigid registration with frame-based techniques. To create a ground truth, the hippocampus was contoured on MR by a trained individual, using RTOG guidelines, before transferring to CT. Then, the images were window and leveled, according to a gaussian approximation of gray matter CT values. An atlas based, center-of-mass method was used to crop the images to 36 consecutive slices. A three channel 2D ResNet with 16,741,358 parameters, built with Keras using the Tensorflow backend was trained with three consecutive, randomly generated slices across four Titan V GPUs with the ADAM optimizer and a dice loss calculated to exclude the background.

Results: The dataset consisted of 136 patients, split 95/14/27 for training, validation and testing, respectively. Evaluation on the test set utilized sliding inference to calculate hippocampal volumes with the median volumetric dice on the right and left hippocampus at 84.7% and 83.2%. Inter-observer dice was calculated from five trained observers with a right and left median of 75.1% and 75.0%, respectively.

Conclusion: Deep learning neural networks for hippocampal segmentation can outperform inter-observer variance on CT images alone. Further work will be to expand the dataset to include patients treated for metastatic disease and identifying if treatment plans differ significantly from human to model generated contours.

Keywords

Segmentation

Taxonomy

IM/TH- image segmentation: CT

Contact Email