Purpose: Manual delineation of target is often time-consuming, and subject to low consistency and poor repeatability. We aimed to develop a deep discriminative model (DDM) to automatically segment rectal tumors with high accuracy on MR images for locally advanced rectal cancer (LARC).
Methods: The proposed framework DDM is consisting of 1) a U-shaped neural network (U-Net) to roughly segment the tumor, and 2) a conditional random field (CRF) to refine the segmentation results. MRI T2 images were pre-processed using N4ITK to correct for intensity inhomogeneities before fed into the DDM framework. In the first stage, the U-Net extracted global context information by the encoder layers, and the decoder layers completed tumor segmentation. The CRF, a specific type of graphical model, used the prediction and the raw image features to estimate the tumor contour through minimization of energy function. The CRF then enhanced boundary constraints of the segmentation results from the U-Net to improve the classification accuracy. The framework was trained for 312 epochs and evaluated with a group of 43 LARC patients. The segmentation results were compared with the ground truth of Gross tumor volume (GTV) contours drawn by an experienced radiation oncologist using evaluation metrics such as Dice similarity coefficient (DSC), etc.
Results: The proposed DDM framework yielded a mean DSC score of 0.883, compared to a mean DSC of 0.852 for the U-Net-based segmentation model. The corrected paired t-test showed significant better performance T score=2.217 (P-value=0.032) of the proposed DDM model comparing with the U-Net model.
Conclusion: We developed a two-step framework for automated segmentation of rectal tumor on T2 MR images. The proposed framework achieved a significantly better segmentation performance than the state-of-art U-Net, which may facilitate MRI-guided treatment planning and plan adaption.
Funding Support, Disclosures, and Conflict of Interest: Key Research and Development Program of Shaanxi (ProgramNo.2022GY-315)
MRI, Segmentation, Treatment Planning