Hierarchical Multi-Stage Feature Fusion Network (HMSFF-Net) for Advanced miRNA Biomarker Classification
N N Jose, Manivannan. Manivannan, Bhukya Krishna, Jangam Ravali, B. Venkatesh, P. Thirumaraiselvan
Abstract
MiRNA profiles are important indicators for cancer diagnosis and treatment. To address these issues, this study offers the Hierarchical Multi-Stage Feature Fusion Network (HMSFF-Net) for reliable miRNA biomarker categorization. HMSFF-Net uses a novel DSRF net with Dual-Stage Resolution Fusion module to capture localized and global features in several miRNA data resolutions. The Self-Adaptive Feature Relevance Mechanism of HMSFF-Net regulates the pass-through of biologically important characteristics to accurately classify cancer. The model uses a Progressive Transfer Learning Framework to fine-tune deeper layers of pre-trained neural networks and gradually adjust earlier levels for improved domain-specific feature extraction. To decrease temporal inconsistency of miRNA datasets [39], HMSFF-Net uses a Temporal Discrepancy Minimization Loss (TDML) method to scale feature extraction to a consistent time. When biomedical applications lack labeled samples, the model's Self-Adaptive Semi-Supervised Learning Strategy refines the pseudo label to enhance learning. Experiments reveal that HMSFF-Net has 98.3% classification accuracy in miRNA benchmark datasets, outperforming previous methods. The model has better noise tolerance, huge data handling, and cancer biomarker pattern universality. The microscopic application of HMSFF-Net's hierarchical analysis, transfer learning, and temporal alignment approach allows oncologists classify miRNA biomarkers with great accuracy.