Related Experiment Video
Updated: Jun 2, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Predicting Metastatic Potential from Histopathological Whole Slide Images of Primary Tumors in Head and Neck Squamous
Tabasum Majeed1, Syed Wajid Aalam2, Tariq Ahmad Masoodi3
1Centre for Artificial Intelligence, Islamic University of Science and Technology, Awantipora, India. tabasum.majeed@iust.ac.in.
Background:
Head and Neck Squamous Cell Carcinoma (HNSCC) ranks as the 6th most prevalent cancer worldwide, imposing a significant burden on global healthcare systems. The progression to metastasis poses a substantial challenge in the clinical management of HNSCC and is a leading cause of patient mortality. Patients diagnosed with distant metastasis (DM) often have limited treatment options and predominantly receive palliative care. Therefore, timely detection of HNSCC patients at risk of metastasis is paramount for proactive intervention and close monitoring, offering the potential for early-stage treatments and enhancing patient survival rates.
Methods:
The pre-processing of Whole Slide Images (WSIs) involved patching, segmentation, background removal, and color normalization. A novel diagnostic framework was introduced using deep learning to predict the likelihood of DM in patients with HNSCC. The proposed model integrates a Self-Attention mechanism with a Residual Network (ResNet), enhancing performance and addressing challenges arising from patch-level variability during pre-processing.
Results:
The proposed model outperformed convolutional neural networks (CNNs), ensemble methods, and classical classifiers such as multilayer perceptrons (MLPs) and logistic regression. It achieved an accuracy of 82.14%, an AUC of 0.80, and an F1 score of 0.82 in predicting metastasis. Robustness and generalizability were further confirmed through stratified 10-fold cross-validation on the TCGA-HNSCC dataset and external validation on the TCGA-ESCA dataset, demonstrating consistent performance across diverse patient subgroups and cancer types.
Conclusions:
The proposed framework showcases potential of using Deep Learning in enhancing clinical decision-making. These findings highlight the use of attention-based models when integrated with deep residual networks offering insights for oncological diagnostics. Future work may focus on improving explainability and interpretability, incorporating multimodal data, and generalizing the approach to other cancer types.