AI-Based Cancer Detection Model for Contrast-Enhanced Mammography

Abstract

Background: The recent development of deep neural network models for the analysis of breast images has been a breakthrough in computer-aided diagnostics (CAD). Contrast-enhanced mammography (CEM) is a recent mammography modality providing anatomical and functional imaging of the breast. Despite the clinical benefits it could bring, only a few research studies have been conducted around deep-learning (DL) based CAD for CEM, especially because the access to large databases is still limited. This study presents the development and evaluation of a CEM-CAD for enhancing lesion detection and breast classification.

Materials & methods: A deep learning enhanced cancer detection model based on a YOLO architecture has been optimized and trained on a large CEM dataset of 1673 patients (7443 images) with biopsy-proven lesions from various hospitals and acquisition systems. The evaluation was conducted using metrics derived from the free receiver operating characteristic (FROC) for the lesion detection and the receiver operating characteristic (ROC) to evaluate the overall breast classification performance. The performances were evaluated for different types of image input and for each patient background parenchymal enhancement (BPE) level.

Results: The optimized model achieved an area under the curve (AUROC) of 0.964 for breast classification. Using both low-energy and recombined image as inputs for the DL model shows greater performance than using only the recombined image. For the lesion detection, the model was able to detect 90% of all cancers with a false positive (non-cancer) rate of 0.128 per image. This study demonstrates a high impact of BPE on classification and detection performance.

Conclusion: The developed CEM CAD outperforms previously published papers and its performance is comparable to radiologist-reported classification and detection capability.

Keywords: breast cancer; cancer detection; computer aided detection; contrast-enhanced mammography; deep learning.

Figure 1

Impact of the image information (a) FROC curves with different input channels and (b) ROC curves with different input channels. The thin lines represent individual models, whereas the ensemble model performance is represented in thick lines. The disks represent radiologists’ diagnostic results from [41,42]. The size of the disk represents the size of the the clinical dataset.

Figure 2

Ground-truth probability for the detected cancer boxes with respect to the cancer detection sensitivity Se${}^l$. The thin lines represent individual models, whereas the ensemble model performance is represented in thick lines.

Figure 3

CEM CAD detection results (blue boxes) and detection scores for two patients’ CEM exams. The first lines correspond to the low energy images and the second line to the recombined images. Red and green boxes represent ground-truth cancers and benign lesions, respectively.

Figure 4

Detection heatmap for different images. Red areas (in (a,b,d)) represents high probability of cancer, while blue areas (in (a,c,e)) represent high probability of benign finding.

Figure 5

FROC curves considering different IoU thresholds.

Figure 6

Example of detection inconsistency with the annotation. The red and blue boxes are, respectively, ground-truth cancer annotation and CAD detected areas. In this cases, the small detected enhancing areas in the left cranio-caudal view are considered as two false positives.

Figure 7

(a) FROC curve and (b) ROC curve for different categories of BPE. Red: low BPE (minimal and moderate grades). Blue: high BPE (moderate and marked grades).