Automated meniscus segmentation and tear detection of knee MRI with a 3D mask-RCNN

3D-Mask RCNN architecture

A 3D-mask RCNN-based model for the detection and segmentation of meniscus tears was developed (Fig. 3). The Mask RCNN is an extension of Faster RCNN, which detects rectangular objects as a regression and classification task [15]. The object mask is produced by another branch of the RCNN [16]. Since the 3D convolution kernels in the network require a lot of memory, we use small image regions (called patches) of size 256 × 256 × 24. To obtain an accurate model for the detection of meniscal tears, these patches are used to train a 3D-Mask RCNN. To reconstruct the entire sagittal PDW sequence for knee MRI, patches from the test set were then applied to the model, and the resulting reconstruction was then reassembled. The meniscal tear probability for MRI was calculated using the projected likelihood of each patch, and the bounding box was utilized to determine the potential meniscal area.

Fig. 3figure 3

Architecture of 3D Mask RCNN

The original mask RCNN model was modified to a 3D version (Fig. 3). In the network, feature pyramids of different scales are extracted using the Residual Network (ResNet)-Feature Pyramid Network (ResNet-FPN) backbone [17, 18]. With FPN (Feature Pyramid Network), top-down and bottom-up features are combined at different levels. ResNet has a 50-layer deep structure in stage 4 (ResNet-50-C4). With ResNet and FPN together, the extraction of features is more accurate and faster. The candidate bounding boxes are generated from input images using a Region Proposal Network (RPN). The extracted feature maps were aligned with the inputs using a quantization-free layer, namely, ROIAlign. Layers such as this reduce misalignments between ROIs and extracted features in RoIPool. The detection branch uses a classifier network and bounding box regression to determine the box probability and location of each proposed ROI. Using a fully connected network (FCN), the mask branch derives probability and location information from the feature maps to predict segmentation masks for each ROI. For all layers, the Rectified Linear Unit (ReLU) is used as the activation function.

3D Mask RCNN Training

Training the 3D-Mask RCNN used He et al.'s initialization strategy [19], which was effective and was developed using Adam optimizer [20]. Only positive ROI (Region of interest) was considered when defining mask loss for meniscal segmentation. Initially, the learning rate was 0.001, which decreased by a factor of 0.5 after every 100 epochs. During training, this learning rate was changed to increase performance and training speed. An object mask branch has been added to RCNN for the prediction of object masks. There are 32 ROIs per mini-batch. The entire network will be trained in the same end-to-end manner as the RPNs.

Faster RCNN weights are initially initialized randomly with a zero-mean Gaussian distribution compared to our framework. Experimentally determined initial learning rates were set at 0.001 for all layers, and decreased by 0.5 after every 50 epochs. There are two mini-batch images per GPU with ROIs of 256 samples each. The loss function is divided into classification loss and bounding box loss, and Smooth L1 regularization is defined in Ren et al [15]. End-to-end training of jointly trained RPN and Faster RCNN is performed to train the entire network.

Performance analysis

The detected target is compared to the real meniscus marked by an experienced radiologist. More specifically, experienced radiologists manually annotated 3D bounding boxes, and true positive (TP) objects were represented by ROIs extracted from locations marked by radiologists. Background or non-meniscal regions were marked as negative cases. For meniscus segmentation, if its intersection with the true meniscus (IOU) is greater than 50%, the test is determined to be TP. The ratio of positive and negative ROI is 3 to 2.

FROC curves are defined as plots of sensitivity against the average number of false positives per joint of the knee. Compute FROC curves by changing the threshold for confidence in object predictions [21].

Evaluation of meniscus tear classification performance of DCNN

The meniscus is divided into medial and lateral meniscus, and bilateral meniscus as a sample. The results of the test set are compared to arthroscopy results, a ROC curve is drawn, and AUC is calculated.

Image segmentation was evaluated using Dice coefficient. Dice is an ensemble similarity measure function commonly used to determine how similar two samples are.

$$Dice\;(A,B) = \frac \right|}}$$

Among them, A is the delineated meniscus area, B is the meniscus area obtained by algorithm segmentation, and the value of Dice is from 0 to 1. The closer the value is to 1, the better the segmentation effect and the more accurate the model.

MRIs of the knees of the test set were independently assessed by two musculoskeletal radiologists who were both full-time and fellowship-trained. (Reader 1: YW, Radiologist with 15 years of experience in musculoskeletal radiology; reader 2: LQQ, Radiologist with 31 years of experience in musculoskeletal radiology) without knowledge of the arthroscopic surgical diagnosis. When there was a discrepancy between them, a consultation must be held to determine the final diagnosis. A state-of-the-art picture archiving system was used to evaluate anonymized data sets once personal or clinical information had been removed under radiological reading room conditions. The readers were blinded to the patients' clinical histories, intraoperative findings, or indications for knee surgery. The classification standard of meniscus injury refers to the Stoller classification method [22]. For each knee, medial and lateral meniscus were evaluated together for the presence or absence of a meniscus tear. As long as one side of the meniscus was torn, the knee was positive for the meniscus tear. According to Stoller's grade and the results of arthroscopic knee surgery, ROC curve was drawn, and AUC was calculated for the radiological evaluation. To compare the DCNN model evaluation with the radiological evaluation without assuming parameters, the resulting output scores were resampled using a bootstrap test. Compare the distribution of the bootstrap variance measure was compared with the observed variance of the measure. Differences between methods were considered significant if the width of the calculated distribution was much smaller than the observed metric. The statistical significance of the performance difference between our DCNN model evaluation and the radiological evaluation was estimated from the ROC curve [23].

留言 (0)

沒有登入
gif