Automated size‐specific dose estimates framework in thoracic CT using convolutional neural network based on U‐Net model

Abstract Purpose This study aimed to develop an automated method that uses a convolutional neural network (CNN) for calculating size‐specific dose estimates (SSDEs) based on the corrected effective diameter (Deff corr) in thoracic computed tomography (CT). Methods Transaxial images obtained from 108 adult patients who underwent non‐contrast thoracic CT scans were analyzed. To calculate the Deff corr according to Mihailidis et al., the average relative electron densities for lung, bone, and other tissues were used to correct the lateral and anterior–posterior dimensions. The CNN architecture based on the U‐Net algorithm was used for automated segmentation of three classes of tissues and the background region to calculate dimensions and Deff corr values. Then, 108 thoracic CT images and generated segmentation masks were used for network training. The water‐equivalent diameter (Dw) was determined according to the American Association of Physicists in Medicine Task Group 220. Linear regression and Bland–Altman analysis were performed to determine the correlations between SSDEDeff corr(automated), SSDEDeff corr(manual), and SSDEDw. Results High agreement was obtained between the manual and automated methods for calculating the Deff corr SSDE. The mean values for the SSDEDeff corr(manual), SSDEDw, and SSDEDeff corr(automated) were 14.3 ± 2.1 mGy, 14.6 ± 2.2 mGy, and 14.5 ± 2.4 mGy, respectively. The U‐Net model was successfully trained and used to accurately predict SSDEs, with results comparable to manual‐labeling results. Conclusion The proposed automated framework using a CNN offers a reliable and efficient solution for determining the Deff corr SSDE in thoracic CT.

with exposure to ionizing radiation while still achieving the required diagnostic information. 4,57][8] In 2011, the American Association of Physicists in Medicine (AAPM) Task Group 204 (TG204) published a patient-specific dose descriptor based on the size of the patient called the size-specific dose estimate (SSDE). 9The SSDE calculation involves multiplying the volumetric CT dose index (CTDI vol , expressed in milligray) by the patient size-dependent conversion factor (f size ).However, the SSDE calculation using geometric parameters according to the TG204 report does not account for the patient's composition or tissue properties that affect radiation attenuation. 10Subsequently, the methodology described in the AAPM TG220 11,12 addresses the patient size estimation limitation in CT scans by using the water-equivalent diameter (D w ) concept to improve the SSDE accuracy.Although several patient dose-monitoring programs can provide the SSDE based on either the effective diameter or D w , the SSDE is still not readily accessible in the patient dose report generated by a CT scanner. 13,14ecently, Mihailidis et al. 15 proposed a method for calculating the patient diameter that uses x-ray attenuation properties to correct the effective diameter for estimating the D w used in the SSDE calculation in chest CT.The results of this method, which only uses the lateral thickness, were comparable to those from the D w -based calculation.Despite those promising findings, that study had some limitations.First, an average relative electron density was scaled only for the lung and other tissues that may affect the measurement accuracy.Second, the sample size in the study was small, thus limiting the generalizability of the results.Finally, the tissues in each segment had to be manually measured for calculating the corrected effective diameter (D eff corr ), which is time-consuming and prone to high inconsistency.These limitations indicate that further research is needed to modify and/or validate the method before it can be widely adopted in clinical practice.Additionally, an automated measurement approach would have advantages over the manual method.
7][18][19] Semantic segmentation, which involves pixel-wise image segmentation, is a well-investigated problem in computer vision and uses deep learning algorithms.The goal of semantic image segmentation is to classify each image pixel into one of a defined set of classes, resulting in the classification of various entities in the image.][22] To the best of our knowledge, no previous studies have described an automated method for calculating the SSDE based on the approach proposed by Mihailidis et al.This study aimed to develop and test a proposed automated framework using a U-Net convolutional neural network (CNN) 21 to determine the SSDE in chest CT with the goal of reducing measurement variability and to assess the correlation between the SSDE determined using the D eff corr and using the D w .

Manual approach
To derive the SSDE, two methods were used to measure each patient image.First, the D w was determined according to the AAPM TG220. 11,12The region of interest (ROI) based on the patient body contour was manually drawn on the axial image to determine the mean CT number and entire area inside the ROI, and the D w value was then calculated using the following equation 11,12 : where, CT(x, y) ROI is the mean CT number in the ROI, and A ROI = ∑ A pixel is the total area in the ROI.Second, the D eff corr , as defined in accordance with Mihailidis  15,23 To improve the calculation accuracy in this study, the thicknesses of waterequivalent tissue in the anterior-posterior (AP) and lateral (LAT) dimensions were determined, and the average relative electron density for bones (ρ e bone = 1.2), 23 in addition to that of lungs and other tissues, was taken into account.Each relative electron tissue density was multiplied by the length of each segment, and then all were summed to give the patient dimension, as calculated according to Equation (2).The D eff corr was calculated as the corrected AP (AP corr eff ) and corrected LAT (LAT corr eff ) dimensions according to Equation (3) 15 : where  s e (j) is the density of the tissue relative to water of the j-segment,l s (j) is the length of the j-segment,with the subscript "s" referring to the tissue segment and superscript N s indicating the number of line segments along the dimension of interest.The illustration of contouring on the transaxial image for the water-equivalent SSDE calculation and the measurements of the AP and LAT dimensions along with the individual distance for the D eff corr calculation are shown in Figure 1.

Automated-SSDE calculation framework
To develop an automated framework for the SSDE calculation, the initial step involved the utilization of a Python-based CNN algorithm.This algorithm was developed using the Keras library and was implemented within a Collaboratory notebook to perform automatic segmentation of tissues in the axial thoracic CT images.The CNN architecture used the U-Net with Resnet34 blocks in the down-sampling path to achieve the desired segmentation. 21Pre-trained models from the ImageNet dataset were used as the starting point to enhance the training process and improve the model's performance. 24,25The CNN architecture that was designed for automated segmentation of the three classes of tissues (lung, bone, and other tissues) and the background region used in calculating dimensions and the D eff corr is shown in Figure 2. The input layer consists of images with a matrix size of 512 × 512.The original CT slices were resized through a down-sampling path, which consisted of five blocks and incorporated 34 convolutional layers from the original ResNet34.Each convolution block was followed by a rectified linear unit (ReLU) activation function and a 2 × 2 max pooling operation with a stride of 2, which reduced the image's spatial dimension.The network included a 2 × 2 up-sampling path, as depicted on the right side of the architecture.The skip connections between the blocks were connected with the down-sampling and up-sampling paths, and the upsampling path was allowed to use the high-resolution feature maps from the early stages of the network without losing information through pooling, thus restoring the image dimensions.A concatenation operation was used to combine the corresponding feature maps to implement the connections. 21The final layer was a 1 × 1convolutional layer, followed by the Softmax activation function, which outputs the four-class mask prediction of the thoracic image, including the background, lung, soft tissue, and bone.
An overview of our automated SSDE calculation framework is presented in Figure 3. DICOM data extracted from the CT images, including the pixel size, CTDI vol , and other attributes, were entered as input data.
To train the CNN model for the D eff corr SSDE in thoracic CT images, 108 input images and their respective segmentation masks were used.The network was trained for 200 epochs, with 20% of the data set aside for testing.The Jaccard Similarity Index, also known as the Intersection over Union (IoU), 26 was used to assess the similarity between the manual label and predicted masks, providing a measure of the model's prediction accuracy.After using the segmented lungs, bone, and other tissue regions to extract the body contours, the major and minor body axes were determined from the contours.The AP and LAT dimensions, calculated using Equation (2), were derived from the major and minor axes and the average relative electron density of the different tissues that had been segmented by the U-Net model.D eff corr was computed according to Equation ( 3), the f-value was computed according to Equation ( 5) and then SSDE was computed according to Equation (4).

Calculation of SSDE based on the D eff corr and D w
The SSDEs (SSDE Deff corr (manual) , SSDE Dw (manual) , and SSDE Deff corr (automated) ), which were derived from the manual D eff corr , manual D w , and automated D eff corr , were calculated according to the following equation [9][10][11][12] : where f is the size-dependent conversion factor to correct patient size based on the D eff corr and D w .The CT dose descriptor in terms of CTDI vol reported by the scanner is the average CTDI vol across all slices of the scan range.7][8] Since a tube voltage of 120 kVp was used, the f value was calculated according to Equations ( 5) and ( 6) as follows 11,12,15 : where D eff corr and D w are patient size specific expressed in centimeters.

Validation
The results of the automated SSDE were compared with those obtained through manual scaling of tissues.
Regression analysis was performed to examine the correlation between the SSDE calculated by the automated method, manual D eff corr method, and D w method.Pearson's correlation test was performed to determine the agreement between the D eff corr SSDE and D w SSDE calculated values.The percentage difference between the automated and manual D eff corr SSDEs was calculated, and the accuracy of the comparison was evaluated by performing the Bland-Altman analysis.

RESULTS
The measurement of all manual D eff corr , automated D eff corr , and D w values were <32 cm in diameter, and all values of the size-dependent conversion factors based on the manual D eff corr (f Deff corr(manual) ), automated D eff corr (f Deff corr(automated) ), and manual D w (f Dw ) were >1.The mean values of CTDI vol , SSDE Deff corr(manual) , SSDE Deff corr(automated) , and SSDE Dw were 8.5 ± 1.7 (range: 4.2−10.5)mGy, 14.3 ± 2.1 mGy (range: 8.5−17.9)mGy, 14.5 ± 2.4 mGy (range: 8.0−17.9)mGy, and 14.6 ± 2.2 (range: 8.6−18.2) mGy, respectively.The results presented in Table 1 show overall agreement between the D eff corr and D w , with both SSDE Deff corr values (obtained using the manual and automated methods) being slightly lower than the SSDE Dw .The SSDE values obtained from the three methods for male and female subjects are presented in Table 2.
The average IoU scores for the segmentation model across various tissue classes in thoracic CT images, including lung, bone, other tissues, and background, were satisfactory, with values of 0.96, 0.86, 0.95, and 0.99, respectively, averaging 0.94. Figure 4 illustrates the results of the CNN-generated segmentation model, which predicted the testing thoracic CT image correctly when juxtaposed with the manually-labeled ground truth comprising three classes of tissues and the background region.Figure 5 shows the box plots of the SSDEs based on the automated method, manual D eff corr , and D w obtained in this study, and Figure 6 shows the SSDE results according to classification by sex.The validation study results show the correlations determined by linear regression (R 2 ), Pearson's correlation coefficient (r),percentage difference,and Bland-Altman analysis of the SSDEs obtained from the three calculation methods (Table 3 and Figures 7a-f ).

DISCUSSION
Accurate estimation of radiation dose in patients undergoing CT scans requires precise measurement of patient size because the dose received depends on the patient factors, pitch factor, scan length, and CT scanner output that is influenced by the parameter

TA B L E 2
The SSDE (mGy) results obtained from the three calculation methods classified by sex.[6]8 Use of SSDE based on the SSDE Deff is a straightforward approach to estimating radiation doses in CT scans.However, in regions with significant x-ray attenuation inhomogeneities, the SSDE Deff may result in an underestimation of patient size and, subsequently, an overestimation of SSDE.

Sex
In this study, we tested a proposed framework of automated SSDE Deff corr that uses deep learning and takes into account the patient's attenuation properties for estimating the radiation dose received by patients during thoracic CT scans.Additionally, we considered the average relative electron density in both the AP and LAT dimensions to further improve the accuracy of the D eff corr , as described by Mihailidis et al. 15 We found that both automated and manual SSDE Deff corr methods had strong correlations with the SSDE Dw , as demonstrated by the results in Figures 7b,c, since we accounted for correction in both the AP and LAT dimensions in a larger patient dataset than that used by Mihailidis et al.The differences in the patient-equivalent diameter between the automated SSDE Deff corr method and SSDE Dw (+13.2%) and between the manual SSDE Deff corr and SSDE Dw (+13.6%) in the present study were similar to the findings reported by Mihailidis et al. in which the discrepancy between the D eff corr and D w was +13.3%.In a study by Juszczyk et al., 27 who used an automated D w to calculate the SSDE and compared the results to those determined by the GE DoseWatch, their mean SSDE was higher than the one in our study.This finding is probably because of the wider range of patient body sizes in Juszczyk et al.'s study, including a maximum chest diameter >32 cm, than those in our study.Juszczyk et al. reported a mean SSDE of 18.8 ± 14.3 (range: 3.4−64.8),whereas in our study, the mean SSDE values calculated by the three methods were smaller.
Considering the sex-dependent results shown in Table 2, all SSDEs calculated by the three methods were higher for males than for females.This finding is probably because males generally have larger body sizes, particularly in the AP dimension, resulting in higher SSDE values (Figure 6).[6]8 However, different CT scanners use different algorithms in their TCM system depending on the acquired projection radiograph.In case of longitudinal mA modulation, the tube current is mainly adjusted on the basis of the change in the AP thickness of the patient along the z-axis.Since the AP diameter of the thoracic region of females differs significantly from those of males, females might benefit from scanning parameters using a lesser dose for this situation.
In this study, an automated method was used for accurate and consistent estimation of SSDE using a CNN and image processing.The U-Net model with ResNet34 was used for image segmentation and achieved generally satisfactory results, with an accuracy of >90% for the four classes of segmented images.The IoU score for bone was slightly lower than for other classes due to its smaller region and more concave shape, which may have affected the D eff corr determination.The accuracy of the model is crucial for SSDE calculations because

F I G U R E 1 (
Left) Manual contouring on the transaxial image for the water-equivalent size-specific dose estimate calculation (DICOM viewer software).(Right) Manual measurement of the anterior-posterior and lateral dimensions along with the individual distance of lung tissue (blue line), other tissues (red line), and bone (green line) for relative electron density correction.DICOM, Digital Imaging and Communications in Medicine.et al.,was calculated.According to the method published by Mihailidis et al., the D eff corr was obtained only in the lateral thickness, and the average relative electron density was scaled for two regions in lung (ρ e lung = 0.3) and other tissues (ρ e tissue = 1.0).

F
I G U R E 2 U−Net architecture with ResNet34 blocks in the down-sampling path for automated segmentation of the three classes of tissues and the background region used in the corrected effective diameter calculation.F I G U R E 3 Flowchart of the automated SSDE based on the corrected effective diameter (D eff corr ) framework.SSDE, size-specific dose estimate.
Abbreviations: Automated method, SSDE based on automatic calculation of the corrected effective diameter; Manual SSDE-D eff corr , SSDE based on the manuallymeasured corrected effective diameter; Manual SSDE-D w , SSDE based on the manually-measured water-equivalent diameter; SSDE, size-specific dose estimate.

F I G U R E 4
Results for automated segmentation using the U−Net model of the three classes of tissues and the background region shown in transaxial thoracic computed tomography images.

TA B L E 3
The results of validation of the SSDE (mGy) obtained from the three calculation methods.: %Diff = [(SSDE−D eff corr(automated) − SSDE−D eff corr(manual) ) / SSDE−D eff corr(manual) ] × 100.Abbreviations: Automated−SSDE, SSDE based on automatic calculation of the corrected effective diameter; Manual SSDE−D eff corr , SSDE based on the manuallymeasured corrected effective diameter; Manual SSDE−D w , SSDE based on the manually-measured water-equivalent diameter; SSDE, size-specific dose estimate.F I G U R E 5 Box plots of the SSDE based on the three calculation methods.SSDE, size-specific dose estimate F I G U R E 6 Box plots of the SSDE based on the three methods for males and females.SSDE, size-specific dose estimate.

F I G U R E 7
Scatter plots and Bland−Altman analysis showing the correlations between the SSDE calculated by the automated method, manually-corrected effective diameter (D eff corr ), and water-equivalent diameter (D w ).SSDE, size-specific dose estimate.
The SSDE (mGy) results obtained from the three calculation methods.
TA B L E 1