Clinical narratives in Non-Hodgkin’s Lymphoma (NHL) diagnosis—pathology reports, radiology flowcytometry, genetics, and other laboratory investigations, and laboratory data—are rich but noisy, limiting reproducibility in downstream analytics. In this study, OncoFusionNet is presented, as a semantic-aware multimodal framework that integrates ontology-guided data validation with deep representation learning for NHL subtype classification. The pipeline consists of four layers: (i) data ingestion from pathology and radiology reports, (ii) a Semantic Quality-Aware Layer (SQAL) that normalizes and structurally parses text while enforcing ontology constraints (ICD-O-3, SNOMED CT, NCIt, RadLex), (iii) cross-modal fusion using transformer-based embeddings with attention and graph integration, and (iv) a calibrated classification head for subtype prediction. We evaluate our system on a corpus of 1,149 de-identified NHL patient records collected between 2010–2024. SQAL improved named-entity recognition accuracy by >15% over dictionary-only baselines and reduced noise from inconsistent markers and staging references. Ontology mapping enabled cross-patient comparability, while multimodal fusion achieved higher macro-F1 and AUROC than unimodal baselines. These results demonstrate that ontology-driven preprocessing coupled with multimodal representation learning substantially enhances data quality and classification performance, laying on the foundation for trustworthy knowledge graphs and clinically reliable decision support.