Prostate cancer remains one of the most prevalent malignancies among men worldwide, where early and accurate diagnosis significantly impacts patient outcomes. Current deep learning approaches for prostate cancer diagnosis face two fundamental limitations: heavy reliance on large-scale annotated multimodal data and insufficient incorporation of clinical expertise embedded in medical knowledge structures. These limitations often result in models that struggle to capture the nuanced relationships between imaging biomarkers, clinical indicators, and pathological states. To address these concerns, a multimodal prostate cancer diagnosis framework integrating medical knowledge graphs with deep learning is proposed, termed Multimodal Prostate Cancer Diagnosis using Knowledge-Enhanced Networks (MP-KDNet). The core architecture employs a knowledge-driven convolutional neural network that fuses heterogeneous data sources, including MRI sequences, clinical laboratory results, and patient history documentation. Through entity disambiguation and knowledge graph embedding techniques, structured clinical knowledge regarding prostate pathology is extracted and transformed into continuous vector representations. These knowledge entity vectors, alongside multimodal feature representations, serve as multi-channel inputs to the convolutional architecture. Multi-scale convolutional kernels capture diagnostic patterns across both fine-grained clinical observations and broader symptom constellations encoded in medical knowledge. Experimental validation on the MIMIC-IV dataset containing 12,847 prostate-related cases demonstrates that MP-KDNet achieves 82.7% diagnostic accuracy, outperforming conventional multimodal fusion, transformer-based imaging analysis, and knowledge graph reasoning baselines. Results confirm that integrating structured clinical expertise with patient-specific multimodal data yields more accurate discrimination among benign prostatic hyperplasia, prostatitis, and adenocarcinoma subtypes than either data-driven or knowledge-driven approaches alone.
Building similarity graph...
Analyzing shared references across papers
Loading...
Xiaodan Zhang
Chao Wang
Journal of Multimedia Information System
Building similarity graph...
Analyzing shared references across papers
Loading...
Zhang et al. (Tue,) studied this question.
www.synapsesocial.com/papers/69d895796c1944d70ce06849 — DOI: https://doi.org/10.33851/jmis.2026.13.1.29