CNN-Based Approaches with Different Tumor Bounding Options for Lymph Node Status Prediction in Breast DCE-MRI

Abstract

Background: The axillary lymph node status (ALNS) is one of the most important prognostic factors in breast cancer (BC) patients, and it is currently evaluated by invasive procedures. Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI), highlights the physiological and morphological characteristics of primary tumor tissue. Deep learning approaches (DL), such as convolutional neural networks (CNNs), are able to autonomously learn the set of features directly from images for a specific task.

Materials and methods: A total of 155 malignant BC lesions evaluated via DCE-MRI were included in the study. For each patient's clinical data, the tumor histological and MRI characteristics and axillary lymph node status (ALNS) were assessed. LNS was considered to be the final label and dichotomized (LN+ (27 patients) vs. LN- (128 patients)). Based on the concept that peritumoral tissue contains valuable information about tumor aggressiveness, in this work, we analyze the contributions of six different tumor bounding options to predict the LNS using a CNN. These bounding boxes include a single fixed-size box (SFB), a single variable-size box (SVB), a single isotropic-size box (SIB), a single lesion variable-size box (SLVB), a single lesion isotropic-size box (SLIB), and a two-dimensional slice (2DS) option. According to the characteristics of the volumes considered as inputs, three different CNNs were investigated: the SFB-NET (for the SFB), the VB-NET (for the SVB, SIB, SLVB, and SLIB), and the 2DS-NET (for the 2DS). All the experiments were run in 10-fold cross-validation. The performance of each CNN was evaluated in terms of accuracy, sensitivity, specificity, the area under the ROC curve (AUC), and Cohen's kappa coefficient (K).

Results: The best accuracy and AUC are obtained by the 2DS-NET (78.63% and 77.86%, respectively). The 2DS-NET also showed the highest specificity, whilst the highest sensibility was attained by the VB-NET based on the SVB and SIB as bounding options.

Conclusion: We have demonstrated that a selective inclusion of the DCE-MRI's peritumoral tissue increases accuracy in the lymph node status prediction in BC patients using CNNs as a DL approach.

Keywords: axillary lymph nodes status (ALNS); bounding box; breast cancer (BC); convolutional neural network (CNN); deep learning (DL).

Figure 1

Tumor lesion segmentation using 3D Slicer software in axial (a), coronal (b), and sagittal (c) MRI projections during the second phase of the post-contrast sequence as demonstrated in a case involving a 56-year-old woman with right invasive ductal breast cancer with unifocal mass-like lesion characterized by spiculated margins and heterogeneous enhancement after contrast medium administration with curve SI/T type III.

Figure 2

Differences in axial projections of the implemented tumor bounding options. (a) In the SFB, a fixed-size 3D bounding box is used; (b) in the SVB option, the smallest 3D bounding box circumscribed to the tumor region is considered, and (c) in the SLVB, the SVB option is applied to each lesion of the patient.

Figure 3

Architectures of the different CNNs used. Panel (a): The SFB-NET is a 3D CNN with three reduction blocks and two fully connected layers. Panel (b): The VB-NET is a 3D CNN with five reduction blocks and two fully connected layers. Panel (c): The 2DS-NET is a 2D CNN with five reduction blocks and two fully connected layers.

Figure 4

Details about the implemented patient-based 10-fold CV. In the i-th iteration, the i-th fold is selected as the test (green), and the previous one is selected as the validation (orange). The remaining folds are included in the training set.

Figure 5

ROC curves of the implemented experiments. The title of each plot suggests the tumor bounding option that is used.

Figure 6

Precision-recall curves of the implemented experiments. The title of each plot suggests the tumor bounding option that is used.

Figure 7

Confusion matrices of the implemented experiments. The title of each matrix suggests the tumor bounding option that is used. In order to have a matrix for each experiment, we merged the predictions of the 10 folds considered as test sets. The colour depends on the number inside the square: the higher the number, the lighter the colour.