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

BACKGROUND: The diagnostic results of magnetic resonance imaging (MRI) are essential references for arthroscopy as an invasive procedure. A deviation between medical imaging diagnosis and arthroscopy results may cause irreversible damage to patients and lead to excessive medical treatment. To improve the accurate diagnosis of meniscus injury, it is urgent to develop auxiliary diagnosis algorithms to improve the accuracy of radiological diagnosis. PURPOSE: This study aims to present a fully automatic 3D deep convolutional neural network (DCNN) for meniscus segmentation and detects arthroscopically proven meniscus tears. MATERIALS AND METHODS: Our institution retrospectively included 533 patients with 546 knees who underwent knee magnetic resonance imaging (MRI) and knee arthroscopy. Sagittal proton density-weighted (PDW) images in MRI of 382 knees were regarded as a training set to train our 3D-Mask RCNN. The remaining data from 164 knees were used to validate the trained network as a test set. The masks were hand-drawn by an experienced radiologist, and the reference standard is arthroscopic surgical reports. The performance statistics included Dice accuracy, sensitivity, specificity, FROC, receiver operating characteristic (ROC) curve analysis, and bootstrap test statistics. The segmentation performance was compared with a 3D-Unet, and the detection performance was compared with radiological evaluation by two experienced musculoskeletal radiologists without knowledge of the arthroscopic surgical diagnosis. RESULTS: Our model produced strong Dice coefficients for sagittal PDW of 0.924, 0.95 sensitivity with 0.823 FPs/knee. 3D-Unet produced a Dice coefficient for sagittal PDW of 0.891, 0.95 sensitivity with 1.355 FPs/knee. The difference in the areas under 3D-Mask-RCNN FROC and 3D-Unet FROC was statistically significant (p = 0.0011) by bootstrap test. Our model detection performance achieved an area under the curve (AUC) value, accuracy, and sensitivity of 0.907, 0.924, 0.941, and 0.785, respectively. Based on the radiological evaluations, the AUC value, accuracy, sensitivity, and specificity were 0.834, 0.835, 0.889, and 0.754, respectively. The difference in the areas between 3D-Mask-RCNN ROC and radiological evaluation ROC was statistically significant (p = 0.0009) by bootstrap test. 3D Mask RCNN significantly outperformed the 3D-Unet and radiological evaluation demonstrated by these results. CONCLUSIONS: 3D-Mask RCNN has demonstrated efficacy and precision for meniscus segmentation and tear detection in knee MRI, which can assist radiologists in improving the accuracy and efficiency of diagnosis. It can also provide effective diagnostic indicators for orthopedic surgeons before arthroscopic surgery and further promote precise treatment.

Study of marrow cell image compression based on quantization threshold.

In this paper, six integer wavelet transforms are compared on marrow cell image energy distribution and their percentage of zero coefficients in every subband for higher performances of marrow cell image compression.

Fractal feature extraction of marrow cell images.

An algorithm based on the cell image center of gravity and its scaleless range (as scale-invariant) is proposed to estimate the fractal dimension of the marrow cell images. Since the texture of a color image contains not only certain statistical similarity on the structure but also the color distributions, two color parameters are extracted from the color images of marrow cells for the fractal dimension calculation. An efficient approach is described for the discrimination between different marrow cells with the combination of the fractal dimension.

A new algorithm for watershed segmentation of cells in marrow.

A new watershed segmentation algorithm of binary images based on the convex property of cells is proposed in this study. A distance map is developed to determine the number of seeds due to iterative erosion. The ultra-erosion aggregate is considered as seeds to separate the cell cluster, and images are grown by dilation from these seeds until all the image pixels are assimilated except the convex boundary. In order to speed up the computation while maintaining a good cell contour, the choice and growing of seeds are under control.