Abstract PS3-04-17: Deep Learning Classification of Inflammatory Breast Cancer Using Multiparametric MRI: A Multi-Sequence Analysis

Abstract Background: The diagnosis of inflammatory breast cancer (IBC) depends upon subjective clinical criteria, often leading to misdiagnosis and delayed treatment. While multiparametric MRI (mpMRI) is sensitive to characteristic IBC features such as global skin thickening, chest wall edema, and non-mass enhancement, its interpretation remains subjective and prone to inter-observer variability. An objective, automated clinical decision support system has the potential to improve patient outcomes enabling earlier and more accurate identification of IBC. Methods: DenseNet121-based deep learning models were developed for the automated classification of IBC using mpMRI sequences from a prospective institutional patient registry. Four separate datasets were analyzed based on sequence availability: T1-weighted without contrast (T1; N=791), diffusion weighted imaging (DWI; N=611), dynamic contrast-enhanced (DCE) imaging (N=484), and T2-weighted with contrast (T2c; N=267). Models for each sequence were evaluated using both midline-cropped single-breast (ipsilateral) images and bilateral full-breast images. For each sequence, stage-aware stratified splitting allocated 80% of patients for training/validation and 20% for testing. Image preprocessing included resizing to 1443 voxels, intensity normalization, and augmentation with random rotations and flips. Binary classification models were trained using 5-fold cross-validation and performance was evaluated using area under the receiver operating characteristic curve (AUC-ROC) on ensemble predictions. Results: Deep learning classification performance demonstrated strong overall discriminative ability across all sequences. DWI achieved the highest ensemble performance (AUC=0.96, mean fold AUC=0.92±0.02), followed by T2c (ensemble AUC=0.96, mean fold AUC=0.89±0.04), T1 (ensemble AUC=0.89, mean fold AUC=0.84±0.04), and DCE (ensemble AUC=0.80, mean fold AUC=0.84±0.02). Stage-specific analysis revealed high accuracy for early-stage disease (cT1-cT3: 92-97%) and classic IBC (cT4d: 94-95% accuracy), but reduced performance in distinguishing cT4b from cT4d cases (40-42% accuracy). Cross-validation demonstrated robust model stability with minimal variance across folds. The ensemble consistently outperformed individual fold averages, with DWI and T2c achieving clinically significant discrimination for overall IBC detection. A bilateral imaging approach did not significantly improve performance compared to analysis of the involved breast alone. Conclusion: Automated deep learning classification of IBC using mpMRI achieved excellent diagnostic performance, with DWI demonstrating superior discriminative ability (AUC=0.96). While the models consistently distinguished classic IBC (cT4d) from early-stage disease, classification of IBC from locally advanced non-IBC with extensive skin involvement and/or edema (cT4b) remains a challenge, reflecting known clinical diagnostic difficulties. Future work will incorporate combined-sequence analysis, probabilistic deep learning for uncertainty quantification, and an expanded cT4b cohort to improve distinction between IBC and non-inflammatory locally advanced breast cancers. These results demonstrate the potential of a deep learning-based clinical decision support system to improve diagnostic accuracy and help guide personalized treatment decisions for patients with IBC. Citation Format: S. Ramezani, H. T. Le-Petross, M. Naser, R. V. Young, T. K. Morris, M. Kai, M. C. Stauder, B. Lim, L. Anthony, MDACC-IBC, C. D. Fuller, W. Woodward, C. R. Goodman. Deep Learning Classification of Inflammatory Breast Cancer Using Multiparametric MRI: A Multi-Sequence Analysis abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-04-17.