Article Text
Abstract
Introduction/Background In cervical cancer diagnosis, dynamic contrast-enhanced magnetic resonance imaging can reflect the access and distribution of blood vessels and tissues and has a certain effect on the evaluation of microvessels in tumors. This work developed and evaluated segmentation potential on DCE-MRI by fully convolutional networks, with aims to provide a clinical auto-delineation tool for subsequent radiotherapy.
Methodology Ninety contrast-enhanced MRI images of patients with cervical cancer were retrospectively enrolled. Sixteen patients did not participate in the model building process in order to verify the generalization ability. Totally 446 slices(512×512) with tumors were annotated by radiologists, among that 358 slices were used for training and 88 slices for testing (figure 1). A symmetric eight-layer deep networks were developed by the nnU-Net framework and the channel dimension was 32, 64, 128, 256, 480, 480, 480, 480, respectively. In addition, the training epoch was 1000 with a random 20% validation set(Initial lr=0.001, optimizer: SGD).
Results Dice similarity coefficient(DSC), 95% Hausdorff distance(95% HD) and average surface distance(ASD) were applied to evaluate the segmentation performance (table 1). The average DSC of all slices was 0.77(median 0.83, maximum 0.95). The average 95% HD was 5.92 mm(median 3.56) and the average ASD was 0.88 mm(median 0.12). 14 of 16 patients’ average DSC exceeded 0.70 and average ASD were less than 1.2 mm. Meanwhile, 10 of 16 patients’ average 95% HD were less than 5 mm.
Conclusion This experimental result indicates that the tumor of cervical cancer on dynamic contrast-enhanced MRI images can be accurately segmented under small sampling, with a great application potential as assistant tool for real-time dynamic delineation. Deeper studies will be conducted by validating this model on a larger sample and enhancing the robustness of the model clinically.