Detection of Lymph Node Invasion Using Digital Holographic Images and Machine Learning

An accurate detection of the spread of colorectal cancer to lymph nodes is a critical factor in determining the patient prognosis and treatment. Standard histological evaluation faces significant challenges that start from the intrinsic fragility of lymph node tissues, making them difficult to process and prepare for analysis, and from the subsequent manual review of slides, demanding extensive reading time from pathologists. There is a need of computer assisted diagnostic companion of the pathologist that provide objectivity and workflow efficacy. To address this, our study introduces an innovative approach that combines label-free imaging with machine learning (ML) for detection of metastatic invasions into lymph nodes. We utilized a Digital Holographic Microscope to obtain quantitative phase images (QPIs) of unstained samples. QPIs were cropped in superpixels that were classified with two supervised machine learning models (coarse and fine models) based on phase-shift parameters extracted from each superpixel. The ML-based procedure classified pixels with 97.91% Sensitivity and 98.62% Specificity into normal and invasion classes, and delineated invasion and areas with high suspicion grade from normal areas in segmented images. Our tool supports the clinicians in identifying malignant areas and focusing their expertise on the tumor boundaries, representing a step forward in computer-assisted cancer diagnosis.