A dual deep neural network for auto-delineation in cervical cancer radiotherapy with clinical validation

Study design and participants

This study was designed to develop deep convolutional neural networks for automatic contouring of the CTV and OARs on cervical cancer CT images. The definition of the CTV was based on the consensus guideline [22], and the contoured OARs were the bowel bag, left and right femoral heads, bladder, and rectum. This study was approved by the Institutional Ethics Review Board of West China Hospital, Sichuan University and waived informed consent.

Between February 2018 and April 2020, the CT images of 203 patients with pathologically proven stage IA1–IB2 cervical cancer who were treated with post-operation radiotherapy were retrospectively collected from three groups led by three senior radiation oncologists in our department of West China Hospital, Sichuan University. The inclusion criteria were: (1) patients with pathologically proven stage IA1–IB2 cervical cancer, (2) who were treated with post-operation radiotherapy, (3) CT scan for positioning, (4) could obtain the CT images. The exclusion criteria were: (1) patients with cervical cancer who are not candidates for radiation therapy, (2) patients with advanced cervical cancer. Specifically, 71, 67, and 65 cases were collected from the three groups led by three senior radiation oncologists, respectively. In the clinical routine, the annotations of the CTV and OARs on CT images in each group were first manually delineated by junior oncologists, and then reviewed and approved by other leading experienced oncologists. Moreover, the anonymized data set consists of 203 patients’ CT images were reconstructed with 3 mm thickness and 0.9 m × 0.9 m in-plane resolution using a GE Revolution ES CT scanner. The patient was in the supine position during CT scans. Before the CT scans, bladder and rectal preparation were performed. We did not use the contrast agent for the bladder filling. Then the 203 patients were randomly divided into three sets, 121 cases of which were randomly selected for training set, 22 cases for validation set, while the remaining 60 cases were testing set in a ratio of approximately 6:1:3. To further evaluate the practical value of the AI-assisted system, 20 additional cases were prospectively recorded and analyzed from August 2020 to November 2020.

Development of deep convolutional neural network models for automatic contouring

In this study, the automatic contouring procedure was implemented in a two-stage method called SegNet. The first stage distinguishes slices of interest from all slices of continuous 3D CT scans. The continuity of these interested slices containing ROIs is essential for the following delineation of ROIs. The second stage is a segmentation task based on the results of stage 1. The proposed SegNet took CT slices as input, and the corresponding automatic contours were calculated as output. We used a dense convolutional network (DenseNet) [27]for the first identification task and a novel encoder-decoder network for the segmentation. The encoder of SegNet consists of residual convolutional blocks [28], and densely connected blocks were used as the backbone of the decoder. SegNet was developed based on UNet by introducing shortcut connections and deeper convolutional layers. The frameworks of SegNet are shown in Additional file 1: Fig. S1. The detailed process and architecture of the two-stage method are given in Additional file 1: Appendix 1.

Quantitative and qualitative evaluation

For objective evaluation, we used sensitivity and area under curve (AUC) to show the recognition accuracy of the first stage identification task. Higher scores represent better continuity of slices. Three widely used quantitative metrics were adopted for the final evaluation of ROI contouring: the volumetric dice similarity coefficient (DSC) [29], the 95% Hausdorff distance (95HD) [30], and the true positive volume fraction (TPVF) [31].

The automatic CTV contours created on the testing set were assessed clinically. A six-point set of objective evaluation criteria was designed following the international guideline [22], as shown in Table 1. The resulting contours of AI models trained with whole multi-group dataset were recorded as SegNet and UNet, and the automatic contours of the same architecture SegNet only trained with a single group data (A, B or C) were called as SegNet(A), SegNet(B), and SegNet(C), respectively. Three radiation oncologists independently graded these automatically segmented CTVs. The score for each case was either 0 or 1: 0- failing the criteria; 1-reaching the criteria. If all the 6 target sites were achieved the criteria in one patient, 6 points were given, and a full score of 60 patients was recorded as 360 points. To avoid bias, each radiation oncologist performed the evaluation of automatic CTVs from each model every other day in a randomized double-blind manner. The final qualitative score for each case was the rounded average score of three experts.

Table 1 A six-point evaluation criterion for the clinical target volume (CTV) delineation in cervical cancer radiotherapyTesting of the AI-assisted system in clinical setting

The proposed SegNet were integrated to develop an AI-assisted system for automatic contouring of ROIs in cervical cancer radiotherapy. The software has been assessed in the Department of Radiotherapy in West China Hospital since August 2020. The detailed running process and workflow of the AI-assisted system was summarized as follows: the proposed two-stage model was integrated into an artificial intelligence (AI)-assisted system that can be used for automatic delineation of the clinical target volume and organs at risk in the cervical cancer radiotherapy treatment. In general, the workflow of the AI-assisted system consists of the following stages:

Step 1: Data transfer. The AI-assisted system has a user interface. Radiologists log into the system with their username and password and then select cases for treatment planning. The software sends a request to retrieve the patient’s CT scans from the PACS system.

Step 2: Automatic delineation. All slices are pre-processed and then used as the inputs to the first stage. Based on the first model’s results, slices likely to contain regions of interest (ROIs) are used as the inputs to the second stage to determine the ROI boundaries. This automatic contouring process does not require any human assistance and eliminates the drawbacks of inter-and intra-delineation variation within the same case.

Step 3: Manual correction. The AI-generated contours are automatically stored in the Ray Station treatment planning system, on which oncologists can directly re-edit the AI-generated ROI boundaries until the plan has been approved.

In the second step in the workflow, SegNet is able to generate the automatic contours of the CTV and OARs for one case in 13.08 s (on a Linux system with 24 GB of RAM and Nvidia RTX 3090 GPUs), and the AI system’s average time to process a case was approximately 2 min, consisting total three-step workflow.

To analyze the potential value of the AI-assisted system, three radiation oncologists with different clinical experience conducted comparative experiments on 20 new patients who were not included in the development cohort. First, each doctor’s manual revision time of the AI-assisted contours was recorded. The doctors’ time to manually contour the same cases from scratch was recorded after 2 weeks. Moreover, all annotations by the three radiation oncologists were finally reviewed by ZP.L. with more than 30 years of clinical experience to evaluate the quality of the radiotherapy planning according to a 2—grade score: 0—secondary revision (the treatment planning should be re-edited to some extent), or 1—minor or no revision (the planning is basically acceptable for clinical radiotherapy treatment). This comparison was developed to assess the potential influence of the AI-assisted results on radiation oncologists’ plan making. If one patient does not need to modify all six target areas, the score is 6 points, and the full score of 60 patients is 360 points.

Statistical analysis

All statistical comparisons were performed using SPSS software. The patient characteristics of age were statistically analyzed, statistical analysis of significant differences in age between the training set, validation set and testing set were performed by the chi-square test. DSC, TPVF and 95HD were computed for all the target regions. The independent sample t-test method was used to compare DSC, TPVF and 95HD between SegNet and UNet. The time used for revising all the CTV and OARs’ contours before radiotherapy planning were recorded as minutes per case. Statistical significance was set at two-tailed P < 0.05.

留言 (0)

沒有登入
gif