Lesion-Aware Generative Artificial Intelligence for Virtual Contrast-Enhanced Mammography in Breast Cancer

TL;DR

The study tackles reducing contrast-agent exposure and radiation in Contrast-Enhanced Spectral Mammography by synthesizing DES-like images from low-energy mammograms using Seg-CycleGAN, a lesion segmentation-guided extension of CycleGAN. The model integrates localized, lesion-focused losses $L'_{cyc}$ and $L'_{id}$ with global losses, guided by segmentation maps $s$, and optimizes with weights $ ext{lambda}_1$, $ ext{lambda}_2$, and a balancing parameter $ abla$ (gamma) to emphasize tumor regions. Evaluated on the CESM@UCBM dataset with 10-fold cross-validation and pretraining on a public CESM dataset, Seg-CycleGAN demonstrates higher PSNR and SSIM than CycleGAN, with qualitative heatmaps confirming improved lesion fidelity while preserving background structure. The work suggests that segmentation-aware generative models can provide safer contrast-free CESM alternatives and potentially augment CAD systems and multimodal diagnostic workflows. Future work could incorporate segmentation maps into discriminators and extend the approach to multimodal explainable AI to enhance trust and clinical interpretability.

Abstract

Contrast-Enhanced Spectral Mammography (CESM) is a dual-energy mammographic technique that improves lesion visibility through the administration of an iodinated contrast agent. It acquires both a low-energy image, comparable to standard mammography, and a high-energy image, which are then combined to produce a dual-energy subtracted image highlighting lesion contrast enhancement. While CESM offers superior diagnostic accuracy compared to standard mammography, its use entails higher radiation exposure and potential side effects associated with the contrast medium. To address these limitations, we propose Seg-CycleGAN, a generative deep learning framework for Virtual Contrast Enhancement in CESM. The model synthesizes high-fidelity dual-energy subtracted images from low-energy images, leveraging lesion segmentation maps to guide the generative process and improve lesion reconstruction. Building upon the standard CycleGAN architecture, Seg-CycleGAN introduces localized loss terms focused on lesion areas, enhancing the synthesis of diagnostically relevant regions. Experiments on the CESM@UCBM dataset demonstrate that Seg-CycleGAN outperforms the baseline in terms of PSNR and SSIM, while maintaining competitive MSE and VIF. Qualitative evaluations further confirm improved lesion fidelity in the generated images. These results suggest that segmentation-aware generative models offer a viable pathway toward contrast-free CESM alternatives.

Figures (4)

• Figure 1: (A) LE image, (B) DES image, (C) Segmentation Map.
• Figure 2: Schematic representation of the proposed method. We propose Seg-CycleGAN, a model based on the CycleGAN architecture designed to generate DES images from LE images. Unlike the standard CycleGAN, Seg-CycleGAN leverages tumor lesion segmentation maps during training to guide the image translation process.
• Figure 3: (A) LE input image, (B) DES target image, (C) CycleGAN output, (D) Seg-CycleGAN with $\gamma=35$ output, (E) Seg-CycleGAN with $\gamma=100$ output.
• Figure 4: Heatmaps generated by comparing the target DES image and synthetic images generated by (A) CycleGAN, (B) Seg-CycleGAN with $\gamma=35$ (C), and Seg-CycleGAN with $\gamma=100$.