A Hybrid U-Net and Vision Transformer Framework with Dimensionality Reduction for Automated Detection of Throat Cancer Metastasis
Year: 2026
Type: article
Abstract: Ensuring the early and accurate detection of lymph node metastasis continues to underpin prognostication and treatment for throat cancer patients. Histopathological whole-slide images (WSIs) are manually analysed by pathologists in clinical practice; however, this process is labour-intensive and requires a significant amount of time, leading to poor reproducibility, inter-observer variability, and delays/inconsistencies in diagnostics. Addressing these limitations, we propose a U-Net segmentation, Vision Transformer (ViT)-based feature extraction, t-Distributed Stochastic Neighbour Embedding (t-SNE) for dimensionality reduction, and a random forest (RF) classification-driven hybrid deep learning framework for fully automated patient-level detection. The U-Net module segments lymph node regions from multimodal scans (MRI and PET), which makes the localisation of somatic lesions clearer. Features extracted from the ViT both represent high and low-spatial-frequency histological signatures and long-range relations that are then mapped via t-SNE to minimise within-cluster distance while maximising between-cluster distances both locally and globally. This then results in robust discrimination between metastatic and non-metastatic lymph nodes with a RF classifier. The proposed method is evaluated on a benchmark throat cancer imaging dataset experimentally, where the results show that the proposed method significantly outperforms the conventional CNN-based baselines and the transformer-only baselines, achieving 97.6%, 96.8% and 95.9% accuracy, sensitivity and specificity, respectively. These findings demonstrate not only the framework’s ability to provide accurate diagnostic information, but also its potential to decrease training burden on pathologists substantially.
Authors: Senthil Jayapal, R. Annamalai
Institution: Amrita Vishwa Vidyapeetham
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $312
OpenAlex ID: W7203632430