Fig 1.
Pipeline of the segmentation and classification process using the SVM classifier.
After obtaining the T1- and T2-weighted images (a-b) from their corresponding sequences, the body region is found using a threshold method on the T2-weighted images (c). Then, the fit procedure is applied to obtain the relaxation maps and other different variables from the received signal model (d-e). These variables are then used to generate the synthetic T1- and synthetic T2-weighted images (f-g). Later, an independent process using the training data is used to generate tissues probabilistic maps, i.e. the probability of a voxel belonging to different tissues (h). At this point, all feature descriptors are normalized (i) to be used by the multi-class SVM classifier (j), which predicts the different classes (k).
Fig 2.
Selected ROIs on a T1-weighted image of the pelvic area.
1) air, 2) fat, 3) muscle, 4) prostate, 5) bone marrow, and 6) bladder.
Fig 3.
Model of a T1 and T2 relaxation signal.
The T1 relaxation curve is simulated with an ideal inversion pulse (α = 2 = 180°) and magnetic field homogenization (S0 = 4095) (a). The T2 signal is also simulated with magnetic homogenization (S0 = 4095). Reported T1 and T2 values at 3 Tesla from the literature were used to generate the signal: fat (385 ms, 121ms), bone marrow (585 ms, 127 ms), muscle (1295 ms, 40 ms), prostate (1700 ms, 74 ms), and bladder (3000 ms, 50 ms) [24,33,34], respectively. TIs and TEs were chosen to cover the complete span of relaxation times of the previous tissues.
Fig 4.
Process by which the synthetic T1-weighted image is computed.
The relaxation times of tissues and the selected inversion times (TI) are used to generate the synthetic T1-weighted images using the model describing a received signal from a perfect inversion pulse.
Fig 5.
Probabilistic shape prior creation.
Starting from training images, the center of mass is localized (a). Then, every voxel in the ROIs is translated into an accumulator grid (b) and the instances of every different tissue in the different bins are counted to create the probabilistic map (c). From the probabilistic map, six-features descriptors are formed: air, fat, muscle, bone marrow, prostate and bladder.
Table 1.
Classification accuracy.
Fig 6.
Relaxation maps and segmentation map of volunteer 1.
Both T1 and T2 relaxations maps show homogeneous time values within the area of each tissue: prostate, muscle, fat, bladder and bone marrow. In the T1 and T2 maps, fat and bone marrow are not visually separable. The same happens in the T2 map for prostate and fat, muscle and bladder. However, the SVM classifier accuracy is rather excellent for prostate = 99.6%, fat = 92.5%, muscle = 99.3%, and bone marrow = 96.7%. Accuracy is lower for bladder (10.4%) and air (45%) mostly because of the small size of these regions.
Fig 7.
Relaxation maps and classification results of volunteer 3.
Even though this volunteer is affected by noncancerous prostatic hyperplasia, the T1 map presents a rather homogeneous range of values in the prostate area (1600–2000 ms). The T1 map also presents a considerable amount of misleading values (≥2000 ms) in the muscle areas. On the T2 map, the muscle areas are well defined (<70ms) and clearly differentiable from the rest of the tissues (prostate, fat, and bone marrow) which seem to merge in a single region. In any case, the SVM accuracy is excellent: prostate = 89.1%, muscle = 98.8%, fat = 100%, and bone marrow = 89.7%. Bladder and air are not present. All features were used to perform the classification.
Fig 8.
Relaxation maps and classification results of volunteer 20.
The T1 relaxation map shows heterogeneous values for the bladder and prostate areas, and homogeneous non-distinctive areas for fat and bone marrow which complicate the classification process. The T2 map only shows two discernible areas: 1) fat-bone marrow-prostate-bladder (≥100ms) and 2) muscle (<100 ms) adding additional complexity to the classification process. Nevertheless, the SVM classifies accurately labeled the different areas: prostate = 89.1%, fat = 100%, muscle = 98.8%, bone marrow = 89.7%, bladder = 71.9% and air was not present. All features were used to perform the classification.