This retrospective analysis was ratified by the institutional Ethics Review Board, and the requirement for informed content was waived.
The database contains 20 BCa patients identified from October 2013 to August 2014. All patients were scanned by a 3.0T MRI scanner (Discovery MR 750; GE Medical Systems) from the Tangdu Hospital. The inclusion criteria were as follows: 1) patients with pathologically-confirmed BCa lesions after operation, 2) the maximal lesion in bladder lumen and its postoperatively pathological findings were archived, and 3) T2W MRI sequence was performed prior to any treatment. The MRI data with poor imaging quality was excluded, which may make the accurate bladder carcinoma segmentation difficult. The T2W sequence (GE Discovery MRI 750 3.0T) was performed to obtain the preoperative bladder images of each patient, and the main parameters of this sequence included repetition time of 2500 ms, echo time of 135 ms, slice thickness of 1 mm, and pixel size of 0.5×0.5 mm2. These patients were then randomly divided into the training set for model development and the validation set for performance assessment, with 10 patients in each set.
Candidate region determination
For T2W MRI sequence of each patient, the CDLS method was used to segment the inner and outer surface of bladder wall . Between the two surfaces, a potential field and a streamline can be generated based on the Laplacian method , in which the thickness of bladder wall is the arc length of a streamline connecting a point on the inner surface and its corresponding point on the outer surface. The bent rate differences between the paired points reflects bladder abnormalities caused by lesions . From these abnormal points, all the voxels on the streamline can constitute a candidate region of BCa, as shown in Fig. 1, which usually contains both cancerous and wall tissues to be separated.
Voxel-feature-based Segmentation of BCa Region
After the previous processing (Fig 1), the bladder wall and the candidate region of each patient can be obtained. In the training set, the cancerous and wall tissues were manually delineated, and 1159 features were extracted from each voxel of them. Then the SVM-RFE method was adopted to first select an optimal subset of features and then distinguish the cancerous and the wall tissues from the voxels . Using the model constructed by the optimal feature subset in the training set, 10 candidate regions obtained from the validation set were used to evaluate the accuracy of proposed method.
1) Volume of interest delineation from training set
In the training set, volumes of interest were previously manually contoured by a radiologist who has 8 years of bladder MRI reading experience. The cancerous VOIs were contoured within the candidate region and away from the bladder wall as much as possible, as described as the yellow contour in Fig. 5. Due to limited voxels and weak boundary between the cancerous and wall tissues, we selected wall voxels near the candidate region as the wall VOIs and tried to keep its number of voxels approximately equal to that of cancerous VOIs.
2) Voxel-based Feature Extraction
Previous studies indicate that intensity and texture features could reflect pathological properties of different tissue types [28, 29], which can be used to distinguish the BCa tissues from wall tissues . Prior to intensity and texture features extraction, the wavelet transform was used to decompose the original image to obtain 16 wavelet images. Thus, a total of 17 images (16 wavelet images and the original image) were used to extract the intensity and texture features.
The intensity features describe the intensity information of the target voxel (x, y, z) and its six neighbor voxels (x − 1, y, z), (x + 1, y, z), (x, y − 1, z), (x, y + 1, z), (x, y, z − 1), (x, y, z + 1). A total of 20 intensity description features were extracted, which contains the intensities of these seven voxels, the mean intensity-values of 3´3 areas centered at the seven voxels respectively, and the intensity-differences between the target voxel and its six neighbor voxels respectively.
In this study, the Leung-Malik (LM) filter bank was used to extract texture features . The LM filter bank consists of 48 filters, which includes 18 first- and 18 second- derivatives of Gaussian-differential filters (6 orientations, 3 scales), 8 Laplacian of Gaussian filters, and 4 Gaussian smoothing filters. The response from 48 filters is taken as 48 texture features for each voxel.
By considering the x, y, and z coordinate values of each voxel as 3 location features, in this study, a total of 1159 features were generated for each voxel, i.e., 3 location features + (20 intensity features + 48 texture features) ´ 17 images.
3) Feature Selection and Classification Using the SVM
Among features obtained from each voxel, some may be correlated and redundant, which may affect the classification performance [30, 32, 33]. In the present study, we used the SVM-RFE method implemented by LIBSVM package , to find the optimal feature subset with the best differentiation performance . After each iteration, the feature with smallest absolute weight was eliminated. Finally, the optimal feature subset was determined using this approach and a 5-fold cross-validation, which contains the first N features with the highest mean accuracy. The classification performance was evaluated by the sensitivity, specificity, accuracy, and area under the curve (AUC) of the receiver operating characteristics (ROC).
4) The segmentation of the cancerous tissues from the candidate regions
Using the optimal feature subset, the SVM prediction model was performed on the validation set. Based on the SVM model, we can obtain the probability value of each voxel belonging to the cancer region. According to the probability, we calculated the “hard” and “soft” boundary to distinguish the cancer and wall regions.
To obtain the “hard boundary”, we used the probability of 0.5 as the threshold, and then segmented the cancer region from the wall region within the candidate region. After classification, a postprocessing, including the maximum connected region (max-region) and void filling, was performed to obtain the continuous boundary. Meanwhile, according to the position of concerned voxel, the “soft boundary” was defined by the probabilities between 0.1 and 0.9, calculated by the SVM prediction model.
5) The accuracy evaluation of proposed segmentation method
In this study, the manual segmentation was treated as the ground truth. The contours of the cancer regions from the validation set were drawn by another two radiologists with 9 years of experience in MRI interpretation. After delineation of each cancer region slice by slice independently, they worked together on the contours according to a consensus reading. The DSC was used to quantitatively evaluate the performance of proposed segmentation method, which can be calculated by DSC(SG, SA) = 2×|SG∩SA|/(|SG|+|SA|), where SG denotes the manual segmentation of radiologists and SA denotes segmentation results from our method.
Measurement of Invasion Depth
Based on the segmented results, the cancer region was excluded from the candidate region. After that, the 3D thickness map of the bladder wall was calculated using the Laplacian method . In the 3D thickness map, the average thickness of bladder wall Tmean was defined by the mean thickness of bladder wall excluding the candidate region to avoid any bias induced by the cancer region, and the minimum thickness of the candidate region Tmin was obtained. In this way, the invasion depth (TID) can be evaluated by Tmean − Tmin, as shown in Fig. 6.
Due to the limitation of tissue biopsies, the exact invasion depth of a BCa could not be obtained. Instead, the value of invasion depth calculated from the proposed segmentation results was compared with that from manual segmentation.
The datasets in this study are currently not available for freely public access owing to the patient privacy concerns, but may be obtainable from the corresponding authors with the reasonable request approved by the institutional review boards.