Deep Learning-Based Recurrence Prediction in HER2-Low Breast Cancer: Comparison of MRI-Alone, Clinicopathologic-Alone, and Combined Models

Abstract

Background/Objectives: To develop a DL-based model predicting recurrence risk in HER2-low breast cancer patients and to compare performance of the MRI-alone, clinicopathologic-alone, and combined models. Methods: We analyzed 453 patients with HER2-low breast cancer who underwent surgery and preoperative breast MRI between May 2018 and April 2022. Patients were randomly assigned to either a training cohort (n = 331) or a test cohort (n = 122). Imaging features were extracted from DCE-MRI and ADC maps, with regions of interest manually annotated by radiologists. Clinicopathological features included tumor size, nodal status, histological grade, and hormone receptor status. Three DL prediction models were developed: a CNN-based MRI-alone model, a clinicopathologic-alone model based on a multi-layer perceptron (MLP) and a combined model integrating CNN-extracted MRI features with clinicopathological data via MLP. Model performance was evaluated using AUC, sensitivity, specificity, and F1-score. Results: The MRI-alone model achieved an AUC of 0.69 (95% CI, 0.68-0.69), with a sensitivity of 37.6% (95% CI, 35.7-39.4), specificity of 87.5% (95% CI, 86.9-88.2), and F1-score of 0.34 (95% CI, 0.33-0.35). The clinicopathologic-alone model yielded the highest AUC of 0.92 (95% CI, 0.92-0.92) and sensitivity of 93.6% (95% CI, 93.4-93.8), but showed the lowest specificity (72.3%, 95% CI, 71.8-72.8) and F1-score of 0.50 (95% CI, 0.49-0.50). The combined model demonstrated the most balanced performance, achieving an AUC of 0.90 (95% CI, 0.89-0.91), sensitivity of 80.0% (95% CI, 78.7-81.3), specificity of 83.2% (95% CI: 82.7-83.6), and the highest F1-score of 0.55 (95% CI, 0.54-0.57). Conclusions: The DL-based model combining MRI and clinicopathological features showed superior performance in predicting recurrence in HER2-low breast cancer. This multimodal approach offers a framework for individualized risk assessment and may aid in refining follow-up strategies.

Keywords: breast neoplasm; deep learning; human epidermal growth factor receptor 2; prediction algorithms; recurrence.

Figure 1

Flow chart of the inclusion and exclusion criteria of the study population.

Figure 2

Overview of the proposed method for recurrence prediction in HER2-low breast cancer. The workflow comprises four major stages: (1) data acquisition, including preoperative MRI (DCE-MRI and DWI sequences) and clinicopathologic information such as tumor size, nodal status, histologic grade, and hormone receptor status; (2) feature extraction from MRI using convolutional neural networks (CNN) and from clinicopathologic data using a multilayer perceptron (MLP); (3) integration of image and non-image features into a combined representation; and (4) binary classification using a deep learning based model to predict recurrence risk. The final output is a recurrence vs. non-recurrence classification.

Figure 3

Proposed architectures. The model was designed to handle both image and non-image data. Non-image features were processed using a Multi-Layer Perceptron (MLP)-based architecture, while image features were extracted using Convolutional Neural Networks (CNNs).

Figure 4

Comparison of prediction performance in test cohorts for the combined, MRI-alone, and clinicopathologic-alone prediction models. The median area under the curve (AUC) was 0.922 for the clinicopathologic-alone model, 0.91 for the combined model, and 0.68 for the MRI-alone model.

Figure 5

Representative case of HER2-low breast cancer with recurrence. The combined model correctly predicted recurrence, whereas the MRI-alone model did not. A 73-year-old woman with an almost entirely fatty breast and minimal background parenchymal enhancement (BPE). Axial contrast-enhanced T1-weighted MR images at (a) 1 min, (b) 2 min, (c) 3 min, (d) 4 min, and (e) 5 min post-contrast administration demonstrate a 1.4 × 1.7 cm irregular-shaped, irregular-margin mass with heterogeneous internal enhancement located in the right lower inner quadrant. The lesion exhibited rapid initial enhancement, followed by a plateau kinetic curve pattern. No axillary lymph node metastasis was identified. This recurrence case was not predicted by the MRI-alone deep learning model but was correctly predicted by the combined model integrating clinicopathological and MRI features.

Figure 6

Representative case of HER2-low breast cancer with recurrence. Both the MRI-alone and combined models correctly predicted recurrence. A 50-year-old woman with heterogeneously dense fibroglandular tissue and moderate background parenchymal enhancement. Axial contrast-enhanced T1-weighted MR images at (a) 1 min, (b) 2 min, (c) 3 min, (d) 4 min, and (e) 5 min post-contrast administration show a 1.8 × 1.6 cm irregular-shaped, spiculated-margin, rim-enhancing mass located at the 12 o’clock position of the left breast. An additional 8 mm enhancing mass is seen posterior to the main lesion, with two other 8 mm enhancing masses laterally that appear interconnected. Although not shown in the images, another enhancing lesion was present in the upper inner quadrant (UIQ), confirming multicentric disease. Both models successfully predicted this recurrence case.