Abstracts

RETURN TO ABSTRACT LISTING


E1683. Prediction of Lymph Node Metastasis Using a Primary Breast Cancer DCE-MRI-Based 4D Convolutional Neural Network
Authors
  1. Dogan Polat; University of Texas Southwestern Medical Center, Department of Radiology
  2. Son Nguyen; University of Texas Southwestern Medical Center, Lyda Hill Department of Bioinformatics
  3. Paniz Karbasi; University of Texas Southwestern Medical Center, Lyda Hill Department of Bioinformatics
  4. Murat Cobanoglu; University of Texas Southwestern Medical Center, Lyda Hill Department of Bioinformatics
  5. Keith Hulsey; University of Texas Southwestern Medical Center, Department of Radiology
  6. Albert Montillo; University of Texas Southwestern Medical Center, Lyda Hill Department of Bioinformatics
  7. Basak Dogan; University of Texas Southwestern Medical Center, Department of Radiology
Objective:
To develop a novel convolutional neural network using MRI images of breast tumors to help predict nodal metastasis in breast cancer patients.

Materials and Methods:
In an IRB approved study, consecutive patients with primary invasive breast cancer who underwent dynamic contrast-enhanced (DCE) breast MRI in our institution between July 2013 and July 2016 were retrospectively reviewed. We collected clinicopathological data (age, ER, HER2, status Ki67 and grade), clinical (cN) and pathologic node (pN) status from electronic health records. Our 4D model included temporal information using 3 acquisition time points after contrast injection. Outputs of 4D convolutional layers were combined with clinicopathologic data to generate hybrid models trained to differentiate cN0 vs cN+ (including cN1, cN2 and cN3) and pN0 vs pN+ (including pN1, pN2 and pN3) disease. A hyperparameter search was used to determine the optimum depth and number of filters for each layer. Classification performance was assessed with the area under curve (AUC) and receiver operating characteristic (ROC) curve with 10-fold nested cross-validation.

Results:
Of 367 patients whose index cancer were analyzed, 227 (61.9%) had cN0 and 140 (38.1%) had cN+ disease. Mean patient age was 51.6 (SD ±11.8)yrs. AUC, sensitivity and specificity of our 4D hybrid model for differentiating cN0 vs cN+ was 0.80(95% CI:0.71-0.87), 84% (95%CI:74%-93%) and 60% (95%CI: 50%-71%) respectively. Of 150 patients eligible for pN analysis, 105(70.0%) were pN0 and 45(30%) were pN+. AUC, sensitivity and specificity of our 4D hybrid model for differentiating pathological N0 vs pN+ was 0.82 (95% CI:0.74-0.89), 87% (95% CI: 80%-94%) and 62% (95% CI:56%-80%).

Conclusion:
A highly sensitive and robust deep learning model which utilizes MRI of breast tumors may help decrease unnecessary axillary diagnostic procedures and facilitate decision-making in breast cancer patients.