Saturday 01 February 2025
Medical image translation models have made significant strides in recent years, enabling the transformation of images between different modalities and institutions. However, evaluating these models is a complex task, as it requires assessing their performance on downstream tasks while also considering the perceptual quality of the translated images.
A new study has proposed a novel metric called RaD (Radiomic Distance), which measures the similarity between two image distributions based on radiomic features extracted from medical images. These features are designed to capture subtle changes in image structure and texture, making them ideal for evaluating the performance of medical image translation models.
The researchers evaluated several state-of-the-art translation models using RaD and other perceptual metrics, including Fr´echet Inception Distance (FID) and Radial Frequency Distance (RadFID). They found that RaD outperformed these traditional metrics in terms of its ability to predict the performance of downstream tasks, such as segmentation and detection.
Moreover, the study demonstrated that RaD can be used for both single-image-level OOD detection and dataset-level OOD detection. In the former case, RaD can identify individual images that are likely out-of-domain, while in the latter case, it can detect whether an entire dataset is out-of-domain relative to a reference dataset.
The researchers also proposed a new metric called nRaDgroup, which normalizes RaD scores to a fixed range and enables more interpretable results. This metric was found to be effective in detecting OOD datasets, with scores ranging from 0 (completely ID) to 1 (completely OOD).
Furthermore, the study explored the use of learned features from ImageNet and RadImageNet for computing RaD scores, which yielded less stable results compared to using radiomic features. This suggests that radiomic features are more robust and suitable for medical image translation tasks.
The findings of this study have significant implications for the development of medical image translation models. By using RaD as a metric, researchers can evaluate the performance of these models in a more comprehensive and accurate manner. Additionally, the ability to detect OOD datasets and individual images can improve the reliability and trustworthiness of medical image analysis pipelines.
In summary, this study has introduced a novel metric called RaD that outperforms traditional perceptual metrics in evaluating the performance of medical image translation models.
Cite this article: “Introducing RaD: A Novel Metric for Evaluating Medical Image Translation Models”, The Science Archive, 2025.
Medical Image Translation, Radiomic Distance, Perceptual Quality, Image Segmentation, Detection, Out-Of-Domain, Image Analysis, Deep Learning, Medical Imaging, Computer Vision
