Add Fréchet Radiomics Distance (FRD-Score) to MONAI Metrics

[RichardObi]

Is your feature request related to a problem? Please describe.
Many existing metrics for evaluating generative, reconstruction, and harmonization pipelines in medical imaging, e.g. those based on neural representations, lack interpretability, are modality-dependent, and often fail to capture clinically meaningful image characteristics. We conducted some recent work proposing and empirically validating the Fréchet Radiomics Distance (FRD) [1,2] showing that, unlike FID, FRD can directly be used for both 2D and 3D images, be conditioned by anatomical masks, is robust to small sample sizes, and correlates more strongly with clinical utility by reflecting anatomically and biomarker-relevant differences. These advantages address key limitations of current metrics (e.g. FID) and support standardized, reproducible comparison of real, synthetic, and cross-domain medical-imaging datasets.

Describe the solution you'd like
We maintain the open-source FRD-Score repository [3] and propose to integrate it into MONAI by refactoring the code into a MONAI-conformant metric class (e.g., FrechetRadiomicsDistance in monai.metrics) that:

• Extracts radiomic feature vectors for two image sets (based on PyRadiomics)
• Computes the Fréchet distance between the resulting feature distributions
• Integrates with MONAI’s existing metric pipeline

Describe alternatives you've considered
We evaluated embedding-based image distribution metrics such as FID and deep-feature distances 1, but these depend on pre-trained networks whose feature spaces may not reflect medical-imaging semantics.

Additional context
[1] Konz, N., Osuala, R., Verma, P., Chen, Y., Gu, H., Dong, H., Chen, Y., Marshall, A., Garrucho, L., Kushibar, K., Lang, D. M., et al. “Fréchet Radiomic Distance (FRD): A Versatile Metric for Comparing Medical Imaging Datasets.” arXiv preprint arXiv:2412.01496. (Comment: Accepted in Medical Image Analysis, subject to minor revisions) https://arxiv.org/abs/2412.01496
[2] Osuala, R., Lang, D. M., Verma, P., Joshi, S., Tsirikoglou, A., Skorupko, G., Kushibar, K., et al.
“Towards Learning Contrast Kinetics with Multi-Condition Latent Diffusion Models.” In: MICCAI 2024, pp. 713–723. Springer Nature Switzerland. arXiv:2403.13890. https://arxiv.org/abs/2403.13890
[3] Official implementation of the Fréchet Radiomic Distance (FRD). https://github.com/RichardObi/frd-score