shows the B-mode image from the Sonix RP scanner. Two pieces of Porcine muscle tissue were prepared and used in this study. Different numbers of HIFU lesions were induced in each of them. For each lesion, the pre-HIFU RF data and B-mode image were acquired. HIFU exposure was on for 40 seconds to induce a thermal lesion. The post-HIFU RF data and B-mode image were acquired at 10 minutes after the HIFU turned off. The region of interest of B-mode image of this frame was separated to estimate the parameters of interest. shows the selected region of interest of reproduced B-mode images of one of the lesions (Tissue #6: Lesion#4) used as the train data corresponding to pre-, during- and post- HIFU.
B-mode image from Sonix RP monitor related to pre-HIFU experiment
Selected region of interest of reproduced B-mode images for pre-, during- and post-HIFU related to Tissue#6:Lesion#4. (a) pre- (b) during- and (c) post- images
A moving hamming window was used to segment each RF data line in such that it divides into a series of segments of length 0.9 mm (60 sample points), assuming ultrasound speed of 1540 m/s in tissue. The tissue was put in a holder, so the position of it was fixed during acquiring pre and post-data for each lesion. This made it possible to compare two registered images for a point by point analysis.
The lateral distance between two neighbouring lesions was approximately 1cm. From lesions created, two lesions (Tissue#6: Lesions#4,6) were used as the train data and the other two as the test data (Tissue#2: Lesions#2,5). In order to choose the lesions to be used as train data, at first we estimated the parameters of Tissue#6: Lesions#4,5,6. We trained the neural network using different mixture of two lesions among these three lesions. Test error of neural network using Lesion#4,6 as train data was 0.0018 which was less than test error of Lesion#4,5 as train data which was 0.0093, and test error of neural network related to Lesion#5,6 as train data which was 0.011. So we used Lesion#4,6 as train data. In the following, first the images of estimated parameters of the train data set related to Tissue#6: Lesion#4 and then the results of test data set related to Tissue#2: Lesion#2 will be presented. Because of large number of images estimated for four lesions, we only present the results of one of the train and one of the test data set to be able to tracking changes of parameters and also compare the result of detecting lesions with and without the neural network. At the end, images produced from the results of neural network will be presented.
Using the corresponding equations and implemented algorithms, all parameters subject to this study were calculated and imaged. shows the corresponding images of post values of estimated attenuation coefficient, IBS, scaling parameter of the Nakagami distribution, frequency dependent scatterer amplitudes and tissue vibration divided by the corresponding pre values. Regarding to colorbar of the images, it is seen that Attenuation coefficient, IBS, scaling parameter of the Nakagami distribution, and frequency dependent scatterer amplitudes in the HIFU lesion site were increased whereas the tissue vibration of the lesion was decreased.
Tissue#6: Lesion#4 (a) Attenuation coefficient, (b) IBS, (c) Scaling parameter of Nakagami distribution, (d) frequency dependent scatterer amplitudes, and (e) tissue vibration images
Besides estimated parameters of Tissue#6: Lesion#4, parameters of Tisue#6: Lesion#6 were estimated and used as train data for neural network. After using these data to train the network two other lesions (Tissue#2: Lesions#2,5) were used for test. shows selected site of reproduced B-mode images of one of the lesions (Tissue#2: Lesion#2) used as test data for pre-, during-and post-HIFU. shows the corresponding images of attenuation coefficient, IBS, scaling parameter of the Nakagami distribution, frequency dependent scatterer amplitudes, and tissue vibration. As in the previous lesion, the attenuation coefficient, IBS, scaling parameter and frequency dependent scatterer amplitudes in the lesion site were increased whereas the tissue vibration of lesion was decreased.
Selected region of interest of reproduced B-mode images regarding to pre-, during- and post-HIFU related to Tissue#2: Lesion#2. (a) pre- (b) during- (c) post- images
Tissue#2:Lesion#2 (a) attenuation coefficient, (b) IBS, (c) scaling parameter of Nakagami distribution, (d) frequency dependent scatterer amplitudes, and (e) tissue vibration images
This information (related to Tissue#2: Lesion#2) besides estimated parameters of Tissue#2: Lesion#5 were used as test data. The values of the neural network for a pixel of interest are either one or zero for being coagulated or not being coagulated, respectively.
For training the network we needed to distinguish a lesion from the surrounding normal tissue, so the post-HIFU B-mode images along with physical examination of tissue cut and the B-mode image registered during HIFU were used to be able to distinguish a lesion site from the normal tissue. For training and testing neural network we used the pixels in the regions of the tissue where the existence of the normal or coagulated tissues was certain. To this end, we first located the centre of the lesion using B-mode images acquired during HIFU and then using the lesion size measured after tissue cut we estimated which pixels were coagulated. These pixels in the centre of the lesion corresponding to the coagulated tissue were then chosen to train and test the MLP neural network algorithm. The training error (mean square error between the network output and the actual target of train data[37
]) of this network was calculated as 9.14 × 10–12
and the test error (mean square error between the network output and the actual target of test data) was calculated as 0.0018.
The trained neural network was used to segment all the sites of interest of data related to tissue samples. We neglected the areas detected by the neural network algorithm as lesion if it's lateral dimension was less than 1.7 mm, We used this value as the FWHM Lateral dimension of HIFU beam focal spot in water was 1.7mm and assumed that as normal tissue. shows the result of neural network for detecting Tissue#2: Lesion#2. Comparing the size of detected lesion (9.6 mm × 8.5 mm) with the actual size of the lesion from physical examination (10.1 mm × 9 mm) shows that we could detect a lesion with the difference of 0.5 mm × 0.5 mm.
Segmentation of normal and coagulated tissue by 5-3-1 neural network
compares the detected size of Lesion#2 of Tissue#2 (Depth × Length) using the mentioned parameters based on each imaged parameter and also using the neural network algorithm developed in this study. It is seen that the neural network can effectively detect the actual lesion size which is 10.1 mm × 9 mm.
Detected size of high intensity focused ultrasound induced-Lesion#2 (Tissue#2) using different methods
The neural network was run using various combinations of the mentioned parameters to find out which combination causes in best result (output of the neural network is more close to actual target) So six scenarios were studied, in each scenario we neglected one of the parameters then we found the sensitivity of the output of neural network to absence of that parameter. The difference of actual area (Depth × Length) of lesion with the estimated area using each scenario divided by the actual area used as indicator of the sensitivity also we substituted the actual area with the estimated area using all parameters combination to determine proportional accuracy. The best parameters combination is selected based on the determined sensitivities. compares the determined sensitivities for detecting Lesion#2 of Tissue#2. This table includes the information of all scenarios’ results and it shows that the best result is obtained when all of the mentioned parameters are used as input features to the neural network. Estimated sensitivities using actual size illustrates that the estimated area using all parameters without Scaling parameter of Nakagami distribution can result in highest error which means this parameter is so effective in lesion detection. Comparing the estimated sensitivities of other scenario shows other parameters help in more accurate detecting of lesion. In the last column of table, each scenario has been compared with the estimation using all parameters and their sensitivities have been compared with the best scenario. It shows that Scaling parameter of Nakagami distribution causes in highest error in estimating lesion size and the vibration is less effective in the neural network.
The determined sensitivity of the output of neural network for detecting Lesion#2 of Tissue#2 using different combination of features
Physical Examination of Tissue
After acquiring all RF data, the tissues were cut into slices and photographed for gross pathology examinations. shows the tissue#6 in which estimated parameters of lesions#4, 6 were used as the neural network train data. shows the tissue#2 in which estimated parameters of lesions#2, 5 were used as the test data.
Tissue cut and folded open, showing HIFU lesions in its middle part (a) measuring length of the lesion. (b) measuring depth of the lesion
Tissue cut and folded open, showing HIFU lesions in its middle part. (a) Tissue cut and folded from middle to measure length of the lesion. (b) Tissue folded close and cut vertically to measure depth of the lesion