MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Multi-Disciplinary Scientific Session
  5. Fully Automated Multi-Organ Segmentation in Abdominal MRI with DenseUnet

Fully Automated Multi-Organ Segmentation in Abdominal MRI with DenseUnet

Y Chen1, 2*, J Xiao1 , L Wang1 , B Sun3 , Z Deng1 , Y Lao1, 4 , N Wang1, 2 , R Saouaf1 , R Tuli1 , D Li1, 2 , W Yang1, 4 , Z Fan1, 2 , (1) Cedars-Sinai Medical Center, Los Angeles, CA, USA (2) University of California Los Angeles, Los Angeles, CA, USA(3) Fujian Medical University Union Hospital, Fuzhou, Fujian, China (4)University of Southern California, Los Angeles, CA, USA

Presentations

(Thursday, 7/18/2019) 10:00 AM - 12:00 PM

Room: 225BCD

Purpose: Accurate and fast contouring of organs-at-risk is highly-desired for online MR-guided adaptive radiotherapy in the abdomen. However, manual delineation is still a common practice and a well-known time-consuming and interobserver variation-prone process. In this work, we developed a deep learning framework for multi-organ segmentation that is fully automated, efficient and accurate.

Methods: Our deep learning network DenseUnet is based on the popular 2D U-net [1] structure, whose input are multiple MRI slices and output are segmentation masks. T1-VIBE (Volumetric Interpolated Breath-hold Examination) images of 61 subjects were collected from our clinical imaging database and split into 49 training and 12 testing sets. For training and evaluation purposes, labels of 10 dose-sensitive organs were manually labelled by two independent radiologists who eventually reached consensus: liver, spleen, pancreas, left/right kidneys, stomach, duodenum, small intestine, spinal cord, and vertebral bodies. All MR images were interpolated into 1.2 mm isotropic resolution (~180 slice/case), and then randomly cropped into 256x160x20 voxels. The manual segmentations were used as the target for the network to optimize. The framework was implemented in Tensorflow [2] and trained for 300k iteration on Nvidia GPUs. The model checkpoints with the best validation accuracy were assessed for performance. Dice coefficient [3] and Jaccard index [4] were used for quantitative analysis of the agreement between model prediction and human label.

Results: Our proposed DenseUnet, inspired by DenseNet [5], has a higher average score (dice: 0.82/Jaccard: 0.72) than Plain-Unet (0.80/0.69). Additionally, our DenseUnet runs 8x faster, which proves that our network structure is more efficient.

Conclusion: Our proposed learning-based automated multi-organ segmentation framework is an efficient and accurate tool allowing fast and precise organ labelling. Further contrast weightings from additional MR sequences are being added to further improve the accuracy for small organs where their contrast to surroundings are inadequate on T1-VIBE.

Keywords

MRI, Segmentation, Radiation Therapy

Taxonomy

IM/TH- image segmentation: MRI

Contact Email

Additional Links