Click here to

Session: Imaging: Mammography and Tomosynthesis [Return to Session]

Deep Learning Method for Volumetric Segmentation of Dense Tissue in Tomosynthesis Using Computer Simulations

J Gomes1, Y Barbosa2, T Do Rego3, T Silva4, T Vent5, A Maidment6, B Barufaldi7*, (1) Universidade Federal Da Paraiba, Joao Pessoa, Brazil,(2) Universidade Federal Da Paraiba, Joao Pessoa, Brazil,(3) Universidade Federal Da Paraiba, Joao Pessoa, Brazil,(4) Universidade Federal Da Paraiba, Joao Pessoa, Brazil,(5) University of Pennsylvania, Philadelphia, PA, (6) University of Pennsylvania, Villanova, PA, (7) University of Pennsylvania, Philadelphia, PA


SU-IePD-TRACK 2-6 (Sunday, 7/25/2021) 3:00 PM - 3:30 PM [Eastern Time (GMT-4)]

Purpose: The use of deep-learning systems in clinical care requires extensive evaluations to develop effective, safe, and acceptable artificial intelligence applications. This study aims to evaluate and optimize nnU-Net models to estimate volumetric breast density (VBD) in digital breast tomosynthesis (DBT) images.

Methods: A virtual clinical trial (VCT) framework was employed to simulate 2,010 anatomical breast models, with distribution of dense compartments varying from 0.1% to 50%. DBT projections were simulated using the acquisition geometry of a clinical tomosynthesis system. The projections were reconstructed using 0.1mm, 0.2mm, and 0.5mm slice spacing. Further, we developed and trained three nnU-Net models using the 51 center-most slices of the DBT reconstructions along with the respective ground-truth tissue labels. For each model, 75% and 25% of ground-truth labels and DBT images were used for training and test, respectively. The effect of the reconstructed slice spacing in the nnU-net models was evaluated using sensitivity (TPR), Jaccard, and Dice metrics.

Results: The accuracy of the nnU-Net varies significantly with smaller reconstructed slice spacing. Images reconstructed with 0.1mm slice spacing resulted in a performance improvement that represents a mean Jaccard increase of up to 41%, 32%, 31%, and 37% (P<0.001), for phantoms simulated using 10%, 20%, 30%, and 50% of dense compartments, respectively. Changes in performance are also noted in the mean Dice and mean TPR metrics.

Conclusion: Computer simulations can optimize DBT acquisition parameters to improve VBD segmentation. Results from simulations with ground truth provide evidence for a better understanding of deep learning outcomes. The nnU-Net performance has been shown to improve the VBD segmentation in DBT using reconstructions with smaller slice spacing. Ultimately, we believe that a plethora of VCT-based strategies will be developed to optimize VBD segmentation.

Funding Support, Disclosures, and Conflict of Interest: Funding Support: BWF IRSA 1016451 and 2020 AAPM Research Seed Grant. Disclosures: A. Maidment: Receives research support from Barco NV, Hologic Inc., and Analogic Inc. Spouse to an employee and stockholder of Real Time Tomography (RTT), LLC. Owner, Daimroc Imaging, LLC Member, RTT Scientific Advisory Board.



    Tomosynthesis, Phantoms, Simulation


    IM- Breast X-Ray Imaging: Phantoms - digital

    Contact Email