Abstract

Mammographic mass retrieval may support case-based com- parison by identifying previously examined lesions with related imag- ing characteristics. Radiomics provides interpretable descriptors of lesion shape, intensity, and texture, whereas deep image embeddings capture complementary visual information. We propose a lightweight radiomics- to-image alignment framework that maps 102 mammographic radiomics features into a frozen DINOv2 ViT-S/14 space. Using 1,204 mass lesions from VinDr-Mammo, a multilayer perceptron was trained with a bidirec- tional CLIP-style contrastive objective using matched radiomics–image pairs as positives and other lesions as negatives. Held-out radiomics queries were projected into DINOv2 space and ranked against a training- image gallery using cosine similarity and unique-study Top-k filtering. The proposed method increased Top-1 radiomics Pearson correlation from 0.390 to 0.679, while BI-RADS differences versus DINOv2 were not statistically significant. In a direct alignment test against 236 held- out lesion-crop embeddings, the corresponding lesion crop was recovered within the Top-10 for 38.6% of queries versus 17.8% for a matched MSE control. An independent ResNet50 evaluator also showed higher Top-10 visual-neighborhood similarity than direct radiomics retrieval (0.663 vs. 0.648, p = 0.0044). These results support contrastive radiomics-to-image alignment as a discriminative cross-modal bridge, while broader clinical benefit remains to be established.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_097.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to Open Review

Open Review Page: https://openreview.net/forum?id=dCdJPjRbvZ

BibTex

@InProceedings{HenMay_Contrastive_MICCAISAT2026,
        author = { Henedak, Mayar T. AND Selim, Sahar AND Elattar, Mustafa},
        title = { { Contrastive Radiomics-to-Image Alignment for Cross-Modal Mammographic Mass Retrieval } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17265},
        month = {pending},
        page = {pending}
}


back to top