Click here to


Are you sure ?

Yes, do it No, cancel

Annotation-Efficient Deep Learning for Free Breathing Proton Magnetic Resonance Imaging Segmentation

D Capaldi1*, F Guo2, T Nano3, O Morin4, L Xing5, G Parraga6, (1) Stanford University, Stanford, CA, (2) Sunnybrook Research Institute, (3) University of California, San Franisco, San Francisco, CA, (4) University of California San Francisco, San Francisco, CA, (5) Stanford Univ School of Medicine, Stanford, CA, (6) Western University, London, ON, CA


(Wednesday, 7/15/2020) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Room: Track 1

Purpose: Deep learning using convolutional-neural-networks (CNNs) has demonstrated ground-breaking performance in various medical imaging applications, including detection, classification, and segmentation. Unfortunately, these data-driven techniques require large and diverse training datasets with expert level annotations that are difficult to obtain and time consuming – hindering the broader use of deep learning for research and clinical applications. Here, our objective is to develop an approach that utilizes deep learning with small datasets and demonstrate the proposed method for free-breathing proton (¹H) magnetic-resonance-imaging (MRI) lung segmentation.

Methods: Free-breathing ¹H MRI was acquired in 20 asthmatics using a balanced steady-state-free-precession-sequence optimized for ventilation imaging (Capaldi et al, Acad Radiol, 2014). Manual segmentation of free-breathing MRI was performed by a single observer (F.G. with 7 yrs experience) on the first 2–3 breathing-cycles (30 slices) for each patient. A volumetric U-net model was implemented with a training/validation/testing dataset split of N=14/2/4. Segmentations generated by the U-net were entered into a continuous-max-flow (CMF) framework subject to total-variation regularization (Guo et al, Med Phys, 2016) to drive the new segmentations (U-net+CMF). Segmentation accuracy was evaluated by comparing U-net and U-net+CMF results to the manual masks using dice-similarity-coefficient (DSC). Shapiro-Wilk test was used to determine normality, and comparison was performed using Wilcoxon paired t-test.

Results: For lung segmentation, there was a DSC of 0.83±0.12 and 0.92±0.03 (p<0.0001) for the U-net and U-net+CMF, respectively. The proposed algorithm results are comparable to previous work using CMF on manual seeding (Guo et al, MRM, 2018). After training (~17mins), the proposed method required ~1min/subject, compared with ~30mins/subject and ~4mins/subject for manual and CMF model with manual seeding, respectively.

Conclusion: We developed and evaluated a pipeline that provides a rapid and accurate method for automated free-breathing MRI lung segmentation as well as demonstrated a novel way to utilize deep learning with small training datasets.

Funding Support, Disclosures, and Conflict of Interest: Dr. Capaldi received funding support from the Natural Sciences and Engineering Research Council (NSERC) of Canada Postdoctoral Fellowship Program.


Not Applicable / None Entered.


IM- MRI : Quantitative Imaging

Contact Email