MDL-CA: A Multimodal Deep Learning Approach with a Cross Attention Mechanism for Accurate Brain Cancer Diagnosis

Sumaira Sarwar¹Saqib Majeed¹Asif Nawaz^1*Ruqia Bibi¹Seung Won Lee^2*

• ¹Pir Mehr Ali Shah Arid Agriculture University, Rawalpindi, Pakistan
• ²Sungkyunkwan University, Jongno-gu, Republic of Korea

In medical imaging and genomics, brain cancer diagnosis remains a critical challenge due to the complex interplay between underlying molecular mechanisms and anatomical abnormalities. Conventional diagnostic methods, such as invasive biopsies, isolated genomic assays, and standalone Magnetic Resonance Imaging (MRI), suffer from limitations including procedural risks, insufficient sensitivity, and incomplete characterization of tumor heterogeneity. These shortcomings often lead to de-layed diagnosis, inaccurate tumor grading, and suboptimal treatment planning. Single-modality data, such as MRI scans or genomic profiles alone, tend to yield suboptimal accuracy and limited biological interpretability in computer-aided diagnostic models. To address these challenges, To overcome these gaps, this study introduces MDL-CA, a novel multimodal deep learning framework that uniquely fuses genomic graph embeddings with MRI features through a cross-attention mechanism to enable biologically informed and highly accurate brain cancer diagnosis. we The proposed MDL-CA, is a Multimodal Deep Learning framework with a Cross-Attention mechanism, which integrates genomic data with brain MRI to improve diagnostic accuracy. The genomic graph embeddings are fused into intermediate MRI feature maps through a cross-modal attention fusion mechanism, enabling the model to capture intricate biological relationships and spatial patterns. This fusion is further optimized to enhance interactions between molecular and anatomical features, resulting in a biologically informed representation. The Entmax sigmoid function is employed to promote sparsity and enhance interpretability in the final classification stage. Data were collected from The Cancer Imaging Archive (TCIA) and The Cancer Genome Atlas (TCGA). Modality-specific feature extraction is performed using a 3D DenseNet for MRI data and a Graph Attention Network (GAT) for genomic data, following rigorous preprocessing. Extensive experiments across four datasets demonstrate that MDL-CA achieves superior performance, with accuracies of 96.22%, 97.14%, 98.46%, and 98.21%, and F1-scores ranging from 95.95% to 98.40%, validating its robustness and generalizability.