TLDR: This research investigates the effectiveness of artificially generated dual-energy subtracted (DES) images, known as virtual DES images, as an alternative to real DES images in contrast-enhanced spectral mammography (CESM) for classifying breast lesions. The goal is to reduce patient radiation exposure by eliminating the need for high-energy image acquisition. The study tested pre-trained U-Net, end-to-end U-Net, and CycleGAN models for generating virtual DES images. While real DES images achieved the highest classification performance (90.35% F1 score), the pre-trained U-Net model performed best among virtual methods (85.59% F1 score). The findings indicate a performance gap, suggesting that while virtual DES images are promising for reducing radiation, further advancements are needed to match the diagnostic quality of real DES images for lesion classification.
Contrast-enhanced spectral mammography (CESM) is a vital imaging technique used in breast cancer diagnosis, particularly effective in cases where standard digital mammography might be inconclusive due to dense breast tissue. This modality produces two main types of images: low-energy (LE) and dual-energy subtracted (DES) images. While both are crucial, the acquisition of high-energy images, necessary for creating DES images, exposes patients to additional radiation. This has led researchers to explore whether artificially generated, or ‘virtual,’ DES images could serve as a viable alternative, potentially reducing patient radiation exposure.
A recent study by Ana C. Perre, Lu´ ıs A. Alexandre, and Lu´ ıs C. Freire delves into this very question, investigating the impact of using virtual DES images on the classification of CESM examinations into malignant and non-malignant categories. Their work, detailed in the paper Are Virtual DES Images a Valid Alternative to the Real Ones?, is the first of its kind to systematically evaluate this specific impact on lesion classification performance.
The Challenge of Radiation and the Promise of Virtual Images
The core motivation behind generating virtual DES images from LE images is to eliminate the need for high-energy image acquisition. This not only streamlines the imaging workflow but, more importantly, reduces the radiation dose to which patients are subjected during a CESM examination. Image-to-image translation techniques, widely used in various medical imaging fields, offer a pathway to achieve this artificial generation.
Methodology: Testing Different Generation Models
The researchers utilized the Categorized Digital Database for Low-Energy and Subtracted Contrast Enhanced Spectral Mammography Images (CDD-CESM), which contains a large collection of LE and corresponding subtracted CESM images. A crucial initial step involved image registration to ensure spatial correspondence between LE and DES images, followed by manual cropping around lesions for both benign and malignant cases.
To generate virtual DES images, the study tested three prominent models: a pre-trained U-Net model, a U-Net model trained end-to-end with the classifier, and a CycleGAN model. These models were chosen based on their proven performance in image-to-image translation within the CESM field. The classification architecture itself was designed to be modular, allowing for experimentation with different virtual DES image generators. The classifier, a RegNetY 1.6GF, received three input images: the low-energy (LE) image, a denoised version of the LE image (LED), and either the real DES image or one of the generated virtual DES images.
Key Findings: Real DES Still Leads, but Virtual Shows Promise
The study’s results, primarily evaluated using the F1 score due to the imbalanced nature of the dataset (more non-malignant than malignant cases), provided clear insights. The classification architecture using real DES images (LE+LED+DES) achieved the highest performance, with an F1 score of 90.35% on the test set. This underscores the significant diagnostic information contained within real DES images, which leverages contrast enhancement to effectively detect malignant features.
Among the virtual DES approaches, the pre-trained U-Net model (LE+LED+U-Net/PT) performed the best, yielding an F1 score of 85.59%. While this is a notable achievement and competitive with baselines that don’t use DES images, it still presents a performance gap compared to using real DES images. The U-Net trained end-to-end (LE+LED+U-Net/EE) showed slightly lower performance at 85.43% F1. The CycleGAN model (LE+LED+CG) performed the least effectively among the virtual DES methods, with an F1 score of 84.30%. This lower performance is attributed to CycleGAN’s tendency to prioritize visual realism, which might lead to the loss of subtle diagnostic features crucial for malignancy detection.
Also Read:
- AI Models Enhance Knee MRI Analysis for Injury Detection
- AI Streamlines 3D Liver Modeling for Surgical Precision
Conclusion: A Step Towards Reduced Radiation
The research confirms that while virtual DES images offer a compelling path to reduce patient radiation exposure, they do not yet fully match the classification quality achieved with real DES images. The discrepancy likely stems from the additional, subtle diagnostic information present in real DES images that current generative models struggle to perfectly replicate. However, the potential for virtual DES image generation is considerable. Future advancements, possibly incorporating domain-specific knowledge, larger datasets, or more sophisticated generative models like diffusion models, could narrow this performance gap. This study provides a crucial foundation for future research aimed at making exclusive reliance on virtual DES images a clinically viable reality.
