TensorFlow-based MobileNetV2 U-Net tumor segmentation and multiparametric MRI radiomics for predicting cervical lymph node metastasis in oral tongue squamous cell carcinoma

TL;DR

This study develops a deep learning model using MRI scans to automatically detect and predict tumor spread in oral tongue cancer patients, aiding surgical planning.

Contribution

A novel TensorFlow-based MobileNetV2 U-Net model for tumor segmentation and metastasis prediction in oral tongue squamous cell carcinoma.

Findings

The MobileNetV2 U-Net achieved a Dice similarity coefficient of 87% on the test set for tumor segmentation.

The fusion nomogram using automated segmentations reached an AUC of 0.93 for predicting lymph node metastasis.

Automated segmentation performed comparably to manual segmentation in predicting metastasis.

Abstract

In oral tongue squamous cell carcinoma (OTSCC) patients, cervical lymph node metastasis profoundly influences prognoses and is central to guiding surgical strategies. Mapping the likelihood of lymph node metastasis across different cervical nodal levels is essential for achieving precise surgical planning.

OTSCC is a prevalent head and neck malignancy. Accurate MRI-based tumor segmentation and prognostic prediction are essential for detecting lymph node metastasis and improving patient survival rates. However, the potential of deep learning techniques has been underexplored in this context.

This retrospective pilot study included 136 OTSCC patients with non-lymph node metastasis and lymph node metastasis who underwent primary and cervical lymph node dissection following baseline MRI. The development of a machine learning approach, incorporating an automatically segmented approach, enables the creation of a model capable of predicting cervical lymph node metastasis based on primary site tumors.

We propose a two-stage OTSCC diagnostic workflow. First, a multiparametric fusion network (MobileNetV2 U-Net) was implemented using TensorFlow to integrate features from contrast-enhanced T1-weighted (CE-T1WI), T2-weighted (T2WI), and T1-weighted (T1WI) MRI sequences and automatically segment primary tumors. Next, radiomic models were constructed to predict cervical lymph node metastasis from these automated segmentations. A fusion nomogram combining radiomic features and clinical data was developed to predict metastasis status. For comparison, a radiomics model using manually segmented CE-T1WIs was also evaluated.

Data from 136 patients (mean age 50.29 ± 12.25 years; 100 men, 36 women) showed that the MobileNetV2 U-Net achieved a Dice similarity coefficient (DSC) of 85% and a mean intersection over union (IoU) of 76% on the training set, and a DSC 87% and an IoU 79% on the independent test set. The fusion nomogram achieved areas under the ROC curves of 0.98 and 0.93 on the training and test sets, respectively, when using the automated segmentation masks. The automated segmentation nomogram performed comparably to the model using manual segmentations for predicting lymph node metastasis.

Our TensorFlow-based MobileNetV2 U-Net provides clinicians with an automated tool to delineate OTSCC tumors and predict cervical lymph node metastasis, potentially aiding personalized surgical planning.

Figures