A Comprehensive Deep Learning System With MGRF Modeling for Predicting Breast Cancer Response to Neoadjuvant Chemotherapy
Ahmed Sharafeldeen, Fatma Taher, Norah Saleh Alghamdi, Eman Alnaghy, Reham Alghandour, Khadiga Mohamed Ali, Sameh Sayed Ahmed Shamaa, Abdelrahman Gamal, Mohammed Asaad Ghazal, Sohail G. Contractor, Ayman S El-Baz · IEEE Access · 2025
Accurate prediction of breast cancer (BC) response to neoadjuvant chemotherapy (NAC) is critical for tailoring treatment strategies and improving patient outcomes. This study introduces a novel deep learning-based framework that integrates multi-parametric magnetic resonance imaging (MRI) (i.e., T1, T2, STIR, and DWI), along with clinical and molecular subtype markers, to classify tumor response into pathological complete response (pCR), partial response (PR), and stable disease (SD). First, tumor regions are delineated across MRI modalities and then modeled using a translation-invariant Markov-Gibbs random field (MGRF) with analytical parameter estimation to capture modality-specific spatial appearance patterns correlated with NAC response. Subsequently, diffusion-weighted MRI is processed to generate apparent diffusion coefficient (ADC) maps, offering quantitative assessment of intratumoral water diffusion and cellularity. Afterward, an adaptive rescaling module (ARM) is proposed to adjust spatial resolution and project volumetric inputs into 2D, enabling compatibility with pretrained convolutional networks. Finally, a customized SEResNet architecture, augmented with Squeeze-and-Excitation (SE) blocks, is introduced to extract modality-specific features which are then fused with clinical and molecular subtypes descriptors for final classification. Evaluated on a cohort of 109 BC patients using leave-one-subject-out (LOSO) cross-validation method, the system achieved an accuracy of 96.33%, a precision of 96.51%, a recall of 96.33%, an F1-score of 96.23%, and a Cohen’s kappa of 94.08%, outperforming its individual components, various pretrained deep learning models, and a state-of-the-art method. These results underscore the value of integrating the appearance model, functional (i.e., ADC) model, adaptive rescaling module, SE blocks, and clinical and molecular subtype markers for the precise prediction of NAC outcomes.