Deep learning models for breast tumor segmentation in DCE-MRI may exhibit disparities in performance across demographic and clinical subgroups, raising concerns about fairness and clinical trustworthiness. In this work, we propose a subgroup-aware in-processing mitigation strategy that integrates divergence-based regularization directly into the training loop. By leveraging interpretable metadata (e.g., age, menopausal status, breast density), we identify subgroups where the model underperforms and assign higher loss weights to these samples in proportion to their divergence from average performance. Our method enables the model to focus training on underrepresented or harder-to-segment subpopulations without requiring external data or post-processing correction. We evaluate our approach on the MAMA-MIA 2025 challenge dataset, demonstrating improvements in both overall segmentation quality and fairness score. Our results highlight the potential of in-processing mitigation as an effective and practical pathway toward equitable medical image segmentation.
Divergence-Aware Training with Automatic Subgroup Mitigation for Breast Tumor Segmentation / Poeta, Eleonora; Vargas, Luisa; Falcetta, Daniele; Marciano', Vincenzo; Pastor, Eliana; Cerquitelli, Tania; Baralis, Elena; Zuluaga, Maria A.. - 16142:(2026), pp. 52-62. (Intervento presentato al convegno Second Deep Breast Workshop, Deep-Breath 2025, Held in Conjunction with MICCAI 2025 (28th International Conference on Medical Image Computing and Computer Assisted Intervention) tenutosi a Daejeon (KOR) nel September 23rd -27th, 2025) [10.1007/978-3-032-05559-0_6].
Divergence-Aware Training with Automatic Subgroup Mitigation for Breast Tumor Segmentation
Eleonora Poeta;Eliana Pastor;Tania Cerquitelli;Elena Baralis;
2026
Abstract
Deep learning models for breast tumor segmentation in DCE-MRI may exhibit disparities in performance across demographic and clinical subgroups, raising concerns about fairness and clinical trustworthiness. In this work, we propose a subgroup-aware in-processing mitigation strategy that integrates divergence-based regularization directly into the training loop. By leveraging interpretable metadata (e.g., age, menopausal status, breast density), we identify subgroups where the model underperforms and assign higher loss weights to these samples in proportion to their divergence from average performance. Our method enables the model to focus training on underrepresented or harder-to-segment subpopulations without requiring external data or post-processing correction. We evaluate our approach on the MAMA-MIA 2025 challenge dataset, demonstrating improvements in both overall segmentation quality and fairness score. Our results highlight the potential of in-processing mitigation as an effective and practical pathway toward equitable medical image segmentation.File | Dimensione | Formato | |
---|---|---|---|
978-3-032-05559-0.pdf
accesso riservato
Tipologia:
2a Post-print versione editoriale / Version of Record
Licenza:
Non Pubblico - Accesso privato/ristretto
Dimensione
1.15 MB
Formato
Adobe PDF
|
1.15 MB | Adobe PDF | Visualizza/Apri Richiedi una copia |
Divergence_Aware_Training_with_Automatic_Subgroup_Mitigation_for_Breast_Tumor_Segmentation.pdf
embargo fino al 21/09/2026
Tipologia:
2. Post-print / Author's Accepted Manuscript
Licenza:
Pubblico - Tutti i diritti riservati
Dimensione
2.03 MB
Formato
Adobe PDF
|
2.03 MB | Adobe PDF | Visualizza/Apri Richiedi una copia |
Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/11583/3003465