3636 - Harnessing Artificial Intelligence to Automate Volumetric Tumor Segmentation of Breast Cancers from Breast Magnetic Resonance Imaging

Purpose/Objective(s): Artificial intelligence (AI) models can help automate delineation of regions-of-interest (ROIs) in clinical imaging studies, such as detecting breast cancers in mammograms, ultrasound and computed tomography studies, but few have achieved three-dimensional (3D) volumetric tumor segmentation on breast magnetic resonance imaging (MRI). The purpose of this study was to compare the performances of two state-of-the-art, industry-standard, Convolutional Neural Network (CNN)-based deep learning (DL) architectures - U-Net and Variational AutoEncoder (VAE)-UNet – in segmenting 3D tumors from breast MRI studies. We hypothesized that, while U-Net can effectively segment volumetric tumors from breast MRI, the addition of the variational auto encoder in VAE-UNet would outperform U-Net in all cases of breast tumor segmentation. We sought to compare the performances of the two architectures on medical imaging data and identify specific cases where each failed or excelled.

Materials/Methods: In this retrospective study, we evaluated pre-treatment T1 post-gadolinium dynamic contrast-enhanced dedicated breast MRI scans from 222 patients with a total of 269 pathology-proven breast tumors. Each of the two DL architectures, U-Net and VAE-UNet, was separately trained to classify tumors at the voxel level across 1000 epochs. The output is a precise delineation of ROIs and segmentation of each tumor for each patient MRI study. The performance of each model was evaluated using 5-fold cross-validation. The performance for each model was measuring using the Dice accuracy metric.

Results: U-Net and VAE-UNet exhibited comparable performances as measured by Dice coefficients. Across the last 400 epochs when training loss plateaued, U-Net achieved a mean Dice coefficient of 65.51% (SD ± 0.068) compared with 66.13% (SD ± 0.067) for VAE-UNet. Both models exhibited monotonic improvement in the Dice coefficient with increasing epoch number, but with a slightly increased slope for VAE-UNet, overcoming U-Net performance after 600 epochs. However, there were distinct cases where VAE-Net outperformed U-Net (Dice score up to 59% better than U-Net). Subsequent analysis indicated these cases occurred when tumor shapes were less spherical (p=0.001).

Conclusion: Our findings suggest that the current industry standard U-Net performs well at general cases of breast cancer volumetric tumor segmentation from 3D MRI studies, whereas VAE-UNet is well suited for tumor segmentation involving complex shapes. Our study results will inform the choice of DL algorithms in research and clinical endeavors that rely on accurate breast cancer tumor segmentation.