Identification of Micrometastasis in Cervical Lymph Nodes - A Machine Learning-Based Approach

Oral squamous cell carcinoma (OSCC) rates have been on the rise globally due to a lack of health care facilities, unaffordable treatment expenses and diagnosis at the advanced stages. Cervical lymph node metastasis is a critically important prognostic factor for OSCC patients. Micrometastatic deposits critically shape clinical staging and treatment choices. Microscopic examination for micrometastases is a slow, labour-intensive, and error-prone process. The use of machine learning on lymph node photomicrographs overcomes manual limitations and enables automated detection of metastatic tissue. This current study employed a convolutional neural network (CNN) algorithm to detect micrometastasis in lymph node sections.

Fifty lymph node archival tissue sections of 30 OSCC cases with modified Papanicolaou (PAP) staining were considered, of which 25 nodes each were metastatic and non-metastatic cases. A comprehensive set of 500 images was acquired using an Olympus Research Microscope (BX53F2), which was equipped with a CCD camera (Jenoptix Gryphax Arktur).

CNN based algorithm was found to be superior compared to the manual method in the detection of micrometastasis. The validation accuracy of the model was 89.36%, classification accuracy of 85%, with a sensitivity of 0.8667 and specificity of 0.8333. Early micrometastasis detection aids tumour upstaging (3 cases), impacting OSCC treatment and prognosis.

The ROC AUC value of 0.9056 indicates a high level of discriminative capability across thresholds, supporting the robustness of the model in detecting micrometastasis. This CNN model has been justified for improved diagnosis and treatment planning of clinically N0 OSCC patients.

The CNN model can function as a supplementary tool to assist pathologic diagnosis, particularly for large-scale populations. CNNs, known for analysing intricate image patterns, can support pathologists by streamlining the identification and evaluation of disease conditions. This support enhances diagnostic efficiency and improves accuracy when managing vast data volumes.