Development and evaluation of a deep learning framework for screening of cervical lymph node metastasis in microscopic images of oral squamous cell carcinoma: A proof of concept study.
Abstract
INTRODUCTION Lymph node (LN) metastasis serves as a critical prognostic indicator in oral squamous cell carcinoma (OSCC), profoundly influencing treatment approaches and survival outcomes. This proof-of-concept study developed and internally evaluated an economical, deep learning (DL)-assisted screening framework for identifying metastatic deposits in microscopic images of LNs from OSCC. The accessibility and operational costs of a WSI pose a major constraint in developing nations and resource-limited environments.
Methods
An in-house dataset of conventional microscopic images (×400 magnification) was prepared from slides of LNs collected from 77 patients (one slide from each patient, 40 metastatic and 37 non-metastatic cases). The 'Training and Validation' subset of this dataset contained 16,426 metastasis-positive and 14,192 non-metastatic image patches of 256 × 256 pixels prepared from the microscopic images and were later augmented and used for training five deep-learning image classification models using the transfer learning approach. Among the five, fine-tuned convolutional neural network model, 'Xception', achieved the best performance metrics without evident overfitting. The in-house dataset also had a strictly separate EVALUATION component/EV subset that contained a total of 1938 metastasis-positive and 1604 metastasis-free microscopic images of LNs of size 1024 × 768 pixels (×400 magnification). Using a custom Python-based framework and the fine-tuned 'Xception' model, a software tool META_DETECT-AI was developed and tested over the whole microscopic images of the EV subset.
Results
The proposed tool demonstrated 97.94% specificity, 99.85% sensitivity, and 98.98% accuracy with visually explainable heatmap generation capabilities over the microscopic images of the EV subset. At the slide level, it showed 100% sensitivity, 71.43% specificity, and 88.24% accuracy.
Conclusion
The developed framework is intended as a decision-support tool for pathologist support and screening and prioritisation of suspicious microscopic images and not as a replacement for slide-level diagnosis or immunohistochemistry. Upon further validation on a larger, heterogeneous, multicentre dataset, this framework may be integrated with augmented reality microscopy/digital pathology workflow.
