E2E-OCT: end-to-end joint learning model using optical coherence tomography images for vocal cord leukoplakia diagnosis.

In our previous study, a home-built handheld OCT system was used to collect OCT images in vocal cord leukoplakia. First, 383 valid OCT images were collected from 12 patients with leukoplakia (including low-risk, high-risk, and malignant types). The best overall accuracy and recall were 92.59% and 93.25% for low-risk, high-risk, and malignant classification, by random forest (RF) model using 5-fold validation. However, low-risk dysplasia with a sensitivity of 87.72% could not meet clinical requirements. Here, we proposed an end-to-end joint learning model using optical coherence tomography (E2E-OCT) images for vocal cord leukoplakia diagnosis. The overall accuracy and recall improved by 5.58% and 4.87%, and especially the sensitivity for low-risk dysplasia improved from 87.72% to 97.48%. Notably, under leave-one-patient-out (LOPO) cross-validation, the model also maintained 96.64% sensitivity for low-risk dysplasia. The ablation experiments and explanation experiments demonstrated the robustness of our model.