Lymph Node Metastasis Estimation of Cervical Cancer using Deep Learning and PET Image

Abstract

Cervical cancer is the fourth most common cancer among women and may involve lymph node metastasis (LNM) as it progresses. The presence or absence of LNM plays a crucial role in staging and determining optimal treatment strategies. However, current non-invasive imaging techniques do not match the diagnostic accuracy of biopsy, leading to clinical issues such as overtreatment or missed therapeutic opportunities. In this study, we constructed a deep learning model based on Swin Transformer to predict LNM using 142 lymph node regions (47 metastasis-positive / 95 metastasis-negative) extracted from PET images. To improve the concentration of relevant diagnostic features, we introduced methods that input central slices of each lymph node instead of all slices, specifically comparing central 3-slice and central 5-slice approaches. Comparison results showed that the model using all slices achieved F1 score of 0.564, the central 5-slice model achieved 0.521, whereas the central 3-slice input model improved F1 score to 0.708. These findings suggest that combining Transformer-based models with focused, limited-slice inputs from central regions can enhance prediction accuracy for LNM in PET image analysis.