Session Item

Locally advanced cervical cancer (LACC) is one of the most frequent malignant tumors among women. LACC is treated partly with ionizing radiation which makes accurate detection and segmentation of cervical tumors a keystone in the treatment of LACC. Manual segmentation of Gross Tumor Volume (GTV) is time-consuming in the radiotherapy (RT) and brachytherapy (BT) workflows and prone to variability and poor reproducibility. Despite its importance, automatic tumor detection and segmentation have rarely been applied to the pelvic region. In this work, we present a fully automatic method to detect and segment the GTVs for LACC in T2-Weighted (T2W) MR images. To the best of our knowledge, it is the first work presenting a comprehensive comparison between different state-of-the-art deep-learning based methods for GTV segmentation in cervical cancer.

For this study, T2W pre-RT images acquired by 18 imaging devices from 82 patients treated for LACC at Gustave Roussy Cancer Campus were gathered. The images were randomly divided into two sub-cohorts of 80% and 20% for respectively model training and model validation. Different recently developed deep neural network models were trained for 2D/3D MRI volume segmentation. Model performance was assessed quantitatively based on Dice Similarity Coefficient (DSC) and 95th Hausdorff distance (HD95). For tackling the overfitting issue and increasing the generalizability of the segmentation, standard data augmentation techniques like random rotation, flip, zoom, contrast adjustment, Gaussian noise, and smoothing were applied to the MRI volumes used in the training set.

The results showed the effectiveness of 2D-segmentation by SegResNet compared with the other state-of-the-art models with average values of DSC = 0.775 and HD95 = 13.9 mm for the segmentation of GTV in LACC. The model achieved the best results of DSC = 0.863 and HD95=8.9.

Cervical tumor segmentation has been performed by different 2D/3D segmentation methods. Segmentation by the 2D SegResNet model achieved the best results. Considering that our model was trained on multi-devices images, the results are promising. Our model can handle additional costs of manual contouring, variability, and poor reproducibility for the detection and segmentation of cervical cancers in T2W MR images. This study can lay solid ground for GTV detection and segmentation in LACC, which can be used for online adaptation of MRI-based RT and BT treatments. In the near future qualitative clinical validation on an independent test should be performed.