ABSTRACT
Significance: Wide-field imaging Mueller polarimetry is an optical imaging technique that has great potential to become a reliable, fast, non-contact in vivo imaging modality for the early detection of, e.g., deceases and tissue structural malformations, such as cervical intraepithelial neoplasia, in both clinical and low-resource settings. On the other hand, machine learning methods have established themselves as a superior solution in image classification and regression tasks. We combine Mueller polarimetry and machine learning, critically assess the data/classification pipeline, investigate the bias arising from training strategies, and demonstrate how higher levels of detection accuracy can be achieved. Aim: We aim to automate/assist with diagnostic segmentation of polarimetric images of uterine cervix specimens. Approach: A comprehensive capture-to-classification pipeline is developed in house. Specimens are acquired and measured with imaging Mueller polarimeter and undergo histopathological classification. Subsequently, a labeled dataset is created within tagged regions of either healthy or neoplastic cervical tissues. Several machine learning methods are trained utilizing different training-test-set-split strategies, and their corresponding accuracies are compared. Results: Our results include robust measurements of model performance with two approaches: a 90:10 training-test-set-split and leave-one-out cross-validation. By comparing the classifier's accuracy directly with the ground truth obtained during histology analysis, we demonstrate how conventionally used shuffled split leads to an over-estimate of true classifier performance (0.964±0.00). The leave-one-out cross-validation, however, leads to more accurate performance (0.812±0.21) with respect to newly obtained samples that were not used to train the models. Conclusions: Combination of Mueller polarimetry and machine learning is a powerful tool for the task of screening for pre-cancerous conditions in cervical tissue sections. Nevertheless, there is a inherent bias with conventional processes that can be addressed using more conservative classifier training approaches. This results in overall improvements of the sensitivity and specificity of the developed techniques for "unseen" images.
