Self-Supervised Learning for Data-Efficient Inflammatory Breast Cancer Classification on MRI

Authors

• Rohan Gajanan Chandraghatgi, Drexel UniversityFollow
• Huong Le-Petross, The University of Texas MD Anderson Cancer CenterFollow
• Mohamed Naser, The University of Texas MD Anderson Cancer CenterFollow
• Megumi Kai, The University of Texas MD Anderson Cancer CenterFollow
• Michael C. Stauder, The University of Texas MD Anderson Cancer CenterFollow
• Bora Lim, The University of Texas MD Anderson Cancer CenterFollow
• MDACC-IBC
• Clifton David Fuller, The University of Texas MD Anderson Cancer CenterFollow
• Wendy A. Woodward, The University of Texas MD Anderson Cancer CenterFollow
• Chelain Rae Goodman, The University of Texas MD Anderson Cancer CenterFollow
• Saleh Ramezani, The University of Texas MD Anderson Cancer CenterFollow

Files

Description

Introduction: 

Inflammatory breast cancer is an aggressive and fatal disease, representing 2-4% of breast cancer cases and 10% of all breast cancer mortalities in the United States. We previously developed a DenseNet121 based deep-learning model distinguishing IBC from LABC across MRI sequences, trained on the largest labeled IBC mpMRI registry (MD Anderson). However, many institutions lack IBC registries for in-house model development. In this work, we aim to assess if self-supervised pretraining on large, unlabeled, publicly available MRI datasets can reduce the number of labels required to achieve similar performance for IBC classification. 

Methods: 

For our pretraining protocol, we compared two approaches: Bootstrap Your Own Latent and Masked Autoencoder, trained exclusively on DCE and T2 Water Dixon MRI sequences from The Cancer Imaging Archive. Each encoder was evaluated at varying labeled-data fractions from the MD Anderson registry. Both frozen and fine-tuned parameters were assessed.  

Results: 

Both BYOL pretraining regimens outperformed non-pretrained with frozen parameters for both DCE (0.78 vs 0.64) and T2 (0.78 vs 0.73) at 100% labels, indicating BYOL’s self-supervised pretraining learned rich representations of IBC features. For fine-tuned single-sequence runs (DCE, T2), from-scratch matched or exceeded all pretraining protocols across label fractions. Notably, frozen BYOL outperformed fine-tuned BYOL at low label fractions (e.g., T2: 0.72 vs 0.57 @ 30% labels), indicating fine-tuning degraded pretrained representations. At 30% labels, frozen BYOL performance matched fine-tuned scratch performance for both DCE (0.69 vs 0.66) and T2 (0.72 vs 0.75), suggesting pretraining could help with an improved fine-tuning regimen. MAE performance was worse than BYOL pretraining and scratch under both frozen and fine-tuned experiments, suggesting MAE learned features are poor. 

Conclusion:   

The results suggest self-supervised pretraining is a promising approach for improving classification of IBC vs LABC under sparse labels.

Program Affiliation

Medical Students Summer in Oncology at Anderson Research (SOAR) Program

Publication Date

7-23-2026

Recommended Citation

Chandraghatgi, Rohan Gajanan; Le-Petross, Huong; Naser, Mohamed; Kai, Megumi; Stauder, Michael C.; Lim, Bora; MDACC-IBC; Fuller, Clifton David; Woodward, Wendy A.; Goodman, Chelain Rae; and Ramezani, Saleh, "Self-Supervised Learning for Data-Efficient Inflammatory Breast Cancer Classification on MRI" (2026). Students Summer in Oncology at Anderson Research (SOAR) 2026. 6.
https://openworks.mdanderson.org/soar26/6