Multimodal fusion artificial intelligence models for pathological diagnosis in early cervical cancer screening

Cervical cancer is a major threat to women's health. Multimodal fusion AI models integrate cytological, histopathological, clinical, and molecular data, potentially improving detection sensitivity, grading accuracy, and screening efficiency. Traditional methods face limitations in subjectivity, efficiency, and primary care resources. This review details the technical architecture, fusion strategies (deep learning, attention mechanisms, cross‑modal alignment), clinical applications, and translational challenges of multimodal fusion AI models. With further refinement and validation, these models may become useful tools for improving cervical cancer screening and prevention.

Introduction

Cervical cancer had ~604,000 new cases and 342,000 deaths globally in 2020. The WHO aims to eliminate cervical cancer by 2030, but China faces a shortage of over 80% of required pathologists. Traditional screening suffers from subjectivity, inefficiency, and poor standardisation. Single‑modal AI models are limited by reliance on a single information dimension. Multimodal fusion technology integrates multiple diagnostic sources, enhancing accuracy, stability, and generalisability. This review synthesises the core technical frameworks, clinical applications, and translational bottlenecks of multimodal fusion AI models.

Limitations of Traditional Screening

• Liquid‑based cytology: High subjectivity (inter‑observer Kappa 0.5–0.7, false‑positive 15–30%, false‑negative 8–15%); 2D morphology only; low efficiency.

• HPV testing: Cannot distinguish transient vs. persistent infection; lacks subtype‑specific risk differentiation; inter‑assay variability.

• Cervical biopsy: Sampling omission 15–30% for small lesions; low diagnostic consistency; poor integration with other data.

Molecular Diagnostic Techniques

• HR‑HPV detection: DNA testing (>95% sensitivity for CIN II) but cannot distinguish integration; E6/E7 mRNA testing (Aptima, RLU/CO ≥1.0) for triage but limited by subtype differentiation.

• DNA methylation: Ultra‑early warning via PAX1/JAM3 (92% sensitivity, 88% specificity, AUC 0.94) or FAM19A4/miR124‑2 (>98% detection).

• ddPCR and NGS: ddPCR for single‑copy quantification; NGS for comprehensive genotyping, integration, and methylation profiling.

Core Technical Framework

Data standardisation: Image modalities (TCT, colposcopy, WSI) undergo colour correction, tiling, and segmentation. Non‑image data (HPV, molecular values, clinical text) are normalised and embedded. Public datasets include Herlev, SIPAKMeD, and CRIC Cervix.

Feature extraction: CNNs (ResNet, DenseNet) and Vision Transformers extract image features; MLP/TabNet and BioBERT extract molecular and text features. Cross‑modal contrastive learning and knowledge distillation align feature spaces.

Fusion strategies:

• Feature‑level fusion (mainstream): Concatenation, channel attention, cross‑modal attention, or GNN‑based fusion.

• Decision‑level fusion: Weighted voting, Bayesian fusion, or stacking for robustness.
  Selection depends on task, data availability, and deployment context.

Training and optimisation: Addresses sample imbalance (Focal Loss), weak annotation, and domain shift (transfer learning, domain‑adversarial networks). Metrics include sensitivity, specificity, AUC, and F1‑score.

Deployment and interpretability: Lightweight deployment via pruning and quantisation for primary care. Grad‑CAM and LIME generate heatmaps and modality contributions to improve clinical trust.

Current Bottlenecks

• Data heterogeneity and lack of standardisation.

• High annotation costs and inconsistent sample quality.

• Trade‑off between fusion efficiency and interpretability.

• Limited clinical validation; "black box" nature reduces clinician trust; privacy concerns hinder data sharing. Primary care infrastructure is insufficient.

Future Directions

• Standardised multicentre datasets with unified protocols.

• Lightweight, interpretable models with XAI and edge computing.

• Full‑process closed‑loop models covering screening‑triage‑diagnosis‑treatment‑follow‑up.

• Deep multi‑omics integration and foundation models for joint image‑molecular‑text interpretation.

• Federated learning for cross‑hospital collaboration without raw data sharing.

• Regulatory frameworks, Class III device approval, and primary care training.

Limitations

Narrative review; non‑systematic. Most studies are retrospective or single‑centre; large‑scale multicentre trials are limited. Economic evaluation is limited. Emerging advances (e.g., large language models) may not be fully covered.

Conclusions

Multimodal fusion AI models integrate cytological, histopathological, clinical, and molecular information, potentially addressing limitations of traditional screening and single‑modal approaches. Evidence suggests improved lesion characterisation and screening efficiency, but multicentre prospective validation is needed before routine implementation.

Full text

https://www.xiahepublishing.com/2996-3427/OnA-2026-00004

The study was recently published in the Oncology Advances.

Oncology Advances is dedicated to improving the diagnosis and treatment of human malignancies, advancing the understanding of molecular mechanisms underlying oncogenesis, and promoting translation from bench to bedside of oncological sciences. The aim of Oncology Advances is to publish peer-reviewed, high-quality articles in all aspects of translational and clinical studies on human cancers, as well as cutting-edge preclinical and clinical research of novel cancer therapies.

Follow us on X: @xiahepublishing

Follow us on LinkedIn: Xia & He Publishing Inc.