System and method for diagnosing pulmonary tuberculosis-pulmonary arterial hypertension by artificial intelligence chest CT (Computed Tomography) imaging

Abstract

The invention discloses a system and method for diagnosing pulmonary tuberculosis-pulmonary arterial hypertension through artificial intelligence chest CT imaging, and the system comprises a data acquisition module which is used for collecting a chest CT image of a patient in the supine position full inspiration period, and exporting the CT image in a DICOM format; the data processing module is used for carrying out full-lung automatic segmentation and lung lobe segmentation VTDLT calculation by using an AI algorithm through AI image analysis software; the quantitative analysis module is used for outputting the damage volume and the number and distribution of the affected lung lobes; the risk prediction module is used for combining the VTDLT with clinical parameters to construct a PH risk prediction model; according to the application, the pulmonary arterial hypertension is screened by combining CT imaging with an artificial intelligence algorithm, an original mode of screening through right heart catheter examination is not changed, a noninvasive, early-stage and accurate PH screening tool is provided for TDL patients, the use requirements of RHC are reduced, and the pain and medical burden of the patients are relieved.

Description

Technical Field

[0001] This invention belongs to the field of medical image processing and artificial intelligence, specifically relating to an artificial intelligence chest CT imaging system and method for diagnosing pulmonary tuberculosis-pulmonary hypertension. Background Technology

[0002] Destruction of lung tissue following tuberculosis can lead to pulmonary hypertension (PH). Its pathological mechanisms may involve a reduction in the cross-sectional area of ​​pulmonary vessels due to lung injury, pulmonary vasoconstriction and remodeling caused by hypoxia, as well as vasculitis and obliterative endotheliitis. The onset of pulmonary hypertension accelerates patient mortality. Currently, the diagnosis of pulmonary hypertension relies on right heart catheterization.

[0003] Currently, the diagnosis of pulmonary tuberculosis (PH) relies on right heart catheterization (RHC), an invasive procedure with high technical barriers and risks, making it difficult to widely implement. Non-invasive methods such as echocardiography have limited accuracy in patients with total lung disease (TDL) due to factors such as emphysema and pleural thickening. Meanwhile, artificial intelligence (AI) has shown great potential in medical image analysis, especially in the quantitative assessment of lung diseases. However, there is currently no AI-based CT image analysis method that can accurately quantify the volume of lung tissue destruction in TDL patients and correlate it with the risk of PH, providing non-invasive and accurate diagnostic support. Therefore, we propose an AI-based chest CT imaging system and method for diagnosing pulmonary tuberculosis-pulmonary hypertension. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based chest CT imaging system and method for diagnosing pulmonary tuberculosis-pulmonary hypertension, providing a non-invasive, early, and accurate PH screening tool for TDL patients, reducing the need for RHC use, and improving diagnostic and treatment efficiency.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence chest CT imaging system for diagnosing pulmonary tuberculosis-pulmonary hypertension, comprising:

[0006] The data acquisition module is used to acquire chest CT images of the patient in a supine position during full inspiration and export the CT images in DICOM format;

[0007] The data processing module uses AI image analysis software and AI algorithms to automatically segment the entire lung and divide lung lobes based on bronchial anatomy; it sets the HU range (-200 to +30) to identify damaged lung tissue and calculates V. TDLT (Quantifying the volume of damaged lung tissue);

[0008] The quantitative analysis module outputs the lesion volume (ml), the number and distribution of affected lung lobes; generates a three-dimensional reconstructed image, marks the lesion area (as shown in the pink area in the attached figure); and provides a correlation analysis chart between the lesion volume and mPAP and PASP.

[0009] Risk prediction module, the risk prediction module will V TDLT Combined with clinical parameters (such as PASP estimated by ultrasound), a PH risk prediction model was constructed.

[0010] The report output module is used to generate structured reports, including text descriptions, charts, and risk warnings.

[0011] Preferably, the acquisition of chest CT images of the patient in a supine position during the full inspiration phase specifically includes:

[0012] Chest CT images of the patient in supine position during full inspiration were acquired using a Siemens 32-slice spiral CT scanner, with a slice thickness of 1 mm, and exported in DICOM format; then imported into AI image analysis software.

[0013] Preferably, the automatic whole-lung segmentation and lobar segmentation specifically include:

[0014] A deep learning model based on U-Net or a similar structure is used for automatic whole-lung segmentation; the lungs are further segmented into lobes and segments based on bronchial anatomy; the HU range (-200 to +30) is set to identify damaged lung tissue, and V is calculated. TDLT (Quantification of the volume of damaged lung tissue).

[0015] Preferably, the automatic whole-lung segmentation specifically includes the following:

[0016] First, identify the high curvature regions in the lung CT image, calculate the maximum principal curvature value (i.e., the maximum normal curvature), and visualize it using color based on its absolute value.

[0017] The two largest surface regions are identified and correspond to the convex regions of the left and right lungs, respectively. An axis-aligned bounding box (AABB) is then constructed for each identified surface region. Subsequently, by checking the bounding boxes to which each surface region belongs, all remaining surface regions are classified as left and right lung regions.

[0018] Preferably, the lung lobe segmentation specifically refers to:

[0019] First, the seed point within the nodule is determined, and then the anatomical structure is modeled using the sub-volume centered on the seed point.

[0020] Then, an AI algorithm is applied to generate a triangular mesh surface based on the scalar field data within the sub-volume;

[0021] The Coendlink shape classification method based on the "visibility" detection process is adopted to determine the surface region of the nodules and identify their corresponding data points through specific seed points.

[0022] Preferably, the step of determining the nodule surface region and identifying its corresponding data points through a specific seed point is as follows: multiple rays are emitted from the seed point s, and the intersection points (p1, p2...p1) of the rays with the triangular mesh generated by MCA are calculated. m );

[0023] Secondly, the first intersection point between the ray and the triangular mesh is defined as a "visible" point, and subsequent intersection points are marked as "invisible". Then, points that are "invisible" to the seed point are filtered out.

[0024] The shape index of visible points is calculated using an AI algorithm, and points classified as convex prisms are removed. Discontinuous regions formed by all visible points are considered as components of the nodule surface.

[0025] Preferably, the calculation of V TDLT (Quantifying the volume of damaged lung tissue) Specifically, this involves automatically calculating the V of each lung segment using an efficient grid simplification algorithm. TDLT Values ​​and air content.

[0026] An AI-based method for diagnosing pulmonary tuberculosis-pulmonary hypertension using chest CT imaging includes the following steps:

[0027] S1. Acquire DICOM data of the patient's chest CT images through the data acquisition module;

[0028] S2, The data processing module uses AI algorithms to automatically segment the whole lung and divide lung lobes;

[0029] S3. The quantitative analysis module automatically identifies and quantifies the volume of destroyed lung tissue (V) based on the Hounsfield unit (HU) threshold (-200HU to +30HU). TDLT );

[0030] S4, Risk Prediction Module will V TDLT Combined with clinical parameters (such as PASP estimated by ultrasound), a PH risk prediction model was constructed.

[0031] S5. The report output module outputs a visual report, including the volume of damage, lung lobe distribution, pH risk level, and diagnostic recommendations.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] (1) This application uses CT imaging combined with artificial intelligence algorithms to screen for pulmonary hypertension, which does not change the original screening method of right heart catheterization. It provides TDL patients with a non-invasive, early and accurate PH screening tool, reduces the need for RHC use, and alleviates patient suffering and medical burden.

[0034] (2) This application greatly reduces the number of invasive examinations, saves medical resources, and improves diagnostic and treatment efficiency, resulting in V TDLT It was significantly correlated with mPAP (r=0.552, p<0.001), and the AUC of the combination with PASP reached 0.917, with a sensitivity of 90% and a specificity of 88.9%, which was significantly better than the single indicator. Attached Figure Description

[0035] Figure 1 This is a three-dimensional model of the damaged lung reconstructed by AI in this invention, where pink represents the damaged area;

[0036] Figure 2 This is a schematic diagram of the automatic lung lobe segmentation of the present invention and an image of the damaged area under CT, wherein the red area is the damaged area under the CT view.

[0037] Figure 3 The scatter plot shows the correlation between the volume of lung damage and the number of lung lobes and hemodynamics, where: A, B, and C represent the volume of lung damage and the number of lung lobes. TDLT Correlation with mPAP, PASP, and PVR; D, E, F: Correlation between the number of damaged lung lobes and mPAP, PASP, and PVR;

[0038] Figure 4 ROC curves (VTDLT, PASP, and combined diagnostic model). Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figures 1-4 This invention provides a technical solution: an artificial intelligence chest CT imaging system for diagnosing pulmonary tuberculosis and pulmonary hypertension, comprising:

[0041] The data acquisition module is used to acquire chest CT images of patients in a supine position during full inspiration and export the CT images in DICOM format;

[0042] A chest CT image of the patient in the supine position during the entire inspiration phase was acquired, specifically as follows:

[0043] Human spinal artery angiography (HRCT) was performed using a 32-slice spiral CT scanner manufactured by Siemens, Germany. The patient was positioned supine and in a state of complete inspiration. The diagnostic parameters were set as follows: tube voltage 120-140 kV; tube current 140 mA; collimator parameters 64 × 0.625 mm; pitch 0.9875; reconstruction slice thickness 1 mm. The original DICOM format images were imported into commercial image analysis software manufactured by Dexin Technology Co., Ltd. of China for processing.

[0044] The data processing module uses AI image analysis software and AI algorithms to automatically segment the entire lung and divide lung lobes based on bronchial anatomy; it sets the HU range (-200 to +30) to identify damaged lung tissue and calculates V. TDLT (Quantifying the volume of damaged lung tissue);

[0045] The automatic whole-lung segmentation is as follows:

[0046] Please see Figure 1 First, high curvature regions in lung CT images are identified, and the maximum principal curvature value (i.e., maximum normal curvature) is calculated and visualized using color based on its absolute value.

[0047] First, these high curvature regions are identified and removed. The system then calculates the maximum principal curvature value (i.e., the maximum normal curvature) and visualizes it using color based on its absolute value.

[0048] The left and right lungs are separated by segmenting the lung surface into several regions. A simple bounding box strategy is used to identify the left and right lungs: since the convex regions of the left and right lungs typically have the largest surface area and enclose the internal concave regions, the two largest surface regions are first identified, corresponding to the convex regions of the left and right lungs, respectively. Next, an axis-aligned bounding box (AABB) is constructed for each identified surface region. Subsequently, by examining the bounding boxes to which each surface region belongs, all remaining surface regions are classified as left and right lung regions. If vertices exist in the overlapping areas of the left and right lung bounding boxes, these vertices are discarded. The identification results of the left and right lung regions are observed from different perspectives. Based on these clustered lung surface data, a closed and smooth lung boundary is finally achieved through implicit surface fitting operations.

[0049] Lung lobe segmentation, specifically:

[0050] A deep learning model based on U-Net or a similar structure is used for automatic whole-lung segmentation; the lungs are further segmented into lobes and segments based on bronchial anatomy; the HU range (-200 to +30) is set to identify damaged lung tissue, and V is calculated. TDLT (Quantifying the volume of damaged lung tissue);

[0051] When performing nodule segmentation, seed points are typically set. Once the seed point within the nodule is determined, the approach first models the anatomical structure of a sub-volume centered on that seed point. A sub-volume is a small cubic region aligned with the coordinate system axes of the CT dataset. To achieve automated nodule segmentation, the sub-volume size needs to be set to completely enclose the largest nodule of clinical interest (e.g., 35×35×35 mm³). Subsequently, the MCA algorithm is applied to generate a triangular mesh surface based on the scalar field data within the sub-volume. High-signal lesions such as pulmonary nodules mainly exhibit three morphological characteristics: spheres (e.g., nodules), cylinders (e.g., blood vessels), and linear or planar structures (e.g., fissures). In medical image analysis, Gaussian curvature (K0) is... G ) and mean curvature (K H As a core indicator for morphological classification, its definition is based on principal curvatures k1 and k2;

[0052]

[0053] By examining the sign combination of Gaussian curvature and mean curvature, a point (or a voxel) is assigned (labeled) as a specific type of shape; the shape classification scheme based on Gaussian distribution / mean curvature performs worse than the scheme proposed by Koenderink and vanDoorn, where the shape index is defined as:

[0054] ;

[0055] Based on Koenderink's shape classification, five types of shapes are defined, including concave ellipsoids, concave cylinders, hyperboloids, convex cylinders, and convex ellipsoids; their shape index values ​​are continuous, namely [-1, -5 / 8), [-5 / 8, -3 / 8), [-3 / 8, 3 / 8), [3 / 8, 5 / 8), and [5 / 8, 1];

[0056] The Coendlink shape classification method, based on a visibility-based detection process, is employed to determine the surface region of nodules and identify their corresponding data points using specific seed points. The specific steps are as follows: First, multiple rays are emitted from the seed point s, and the intersection points (p1, p2...p...) of these rays with the triangular mesh generated by the MCA are calculated. m Secondly, the first intersection point between the ray and the triangular mesh is defined as a "visible" point, and subsequent intersection points are marked as "invisible". Then, points that are "invisible" to the seed point are filtered out. Thirdly, according to the formula... The shape index of the visible points is calculated, and points classified as convex prisms are discarded. Ultimately, the discontinuous region formed by all visible points is considered part of the nodule surface.

[0057] The quantitative analysis module outputs the lesion volume (ml), the number and distribution of affected lung lobes; generates a three-dimensional reconstructed image, marks the lesion area (as shown in the pink area in the attached figure); and provides a correlation analysis chart between the lesion volume and mPAP and PASP.

[0058] Risk prediction module, the risk prediction module will V TDLT Combined with clinical parameters (such as PASP estimated by ultrasound), a PH risk prediction model was constructed.

[0059] The report output module is used to generate structured reports, including text descriptions, charts, and risk warnings.

[0060] Preferably, the calculation of V TDLT Specifically, it involves automatically calculating the V value for each lung segment using an efficient grid simplification algorithm. TDLT Values ​​and air content:

[0061] The preprocessed CT lung parenchyma data is converted into tetrahedral / voxel meshes, and the complexity is reduced by mesh simplification algorithms (to avoid computational redundancy) while preserving the anatomical features of lung segments.

[0062] 1. Calculate the "error cost" of each grid cell (e.g., the quadratic error matrix in QEM, reflecting the impact of cell deletion on the overall shape); 2. Sort the cells by error cost from smallest to largest, and prioritize deleting cells with the smallest error (e.g., merging adjacent low-gradient cells); 3. After deleting a cell, update the error cost of its adjacent cells and check if the simplification error threshold is met; 4. Repeat steps 2-3 until the number of grid cells is reduced to the target complexity (e.g., from 10). 7 Reduced to 10 5 (balancing efficiency and accuracy).

[0063] Extract all simplified mesh units within the lung segment and exclude non-lung tissue units (such as blood vessels and bronchi, filtered by CT value > -500HU).

[0064] The total volume of this segment is V total : Sum the volumes of all lung segment units (voxel mesh: unit volume = voxel side length) 3 Tetrahedral mesh: element volume = 1 / 6 × base area × height);

[0065] Dosage target volume V D : Accumulate the unit volume of "dose value ≥ preset threshold D" within this segment (D needs to be set according to clinical needs, such as D=20Gy or 30Gy commonly used in radiotherapy);

[0066] The average density of this section is T avg The average density is calculated by weighting the unit volume, using the following formula:

[0067] ;

[0068] In the formula: Let be the density of the i-th unit. Let be the volume of the i-th unit.

[0069] V TDLT Value calculation:

[0070]

[0071] An AI-based method for diagnosing pulmonary tuberculosis-pulmonary hypertension using chest CT imaging includes the following steps:

[0072] Step 1: Acquire DICOM data of the patient's chest CT images through the data acquisition module;

[0073] Step 2: The data processing module uses AI algorithms to automatically segment the whole lung and divide the lung lobes.

[0074] Step 3: The quantitative analysis module automatically identifies and quantifies the volume of destroyed lung tissue (V) based on the Hounsfield unit (HU) threshold (-200HU to +30HU). TDLT );

[0075] Step 4: The risk prediction module will V TDLT Combined with clinical parameters (such as PASP estimated by ultrasound), a PH risk prediction model was constructed.

[0076] Step 5: The report output module outputs a visual report, including the volume of damage, distribution of lung lobes, pH risk level, and diagnostic recommendations.

[0077] The specific implementation method is as follows:

[0078] Of the 161 patients, 55 had active pulmonary tuberculosis (34.16%) and 106 had inactive pulmonary tuberculosis (65.84%). Compared with patients with inactive pulmonary tuberculosis, patients with active pulmonary tuberculosis had a significantly higher probability of experiencing symptoms such as cough, increased sputum production, wheezing, and hemoptysis (P<0.05). The proportion of patients with active pulmonary tuberculosis who developed pulmonary hypertension was also significantly higher than that in the inactive pulmonary tuberculosis group (61.82% vs. 22.64%, P<0.01).

[0079] In the initial analysis, 55 patients with TDL and pulmonary hypertension underwent further screening. Four patients were excluded due to other causes of pulmonary hypertension, resulting in a final study cohort of 51 TDL patients. Their baseline characteristics are detailed in Table 1. Patients were divided into two groups based on their PH status: TDL with PH (n=31) and TDL without PH (n=20). Baseline characteristics are detailed in Table 1. No significant differences were observed between the two groups in terms of sex, age, vital signs, or comorbidities. The NT-proBNP level was significantly higher in the TDL group with PH than in the non-PH group (P<0.05). The following pulmonary function parameters showed significant inter-group differences: forced vital capacity (FVC, predicted value%), forced expiratory volume in one second (FEV1, predicted value%), FEV1 / FVC ratio, and DLCO (P<0.05). Right ventricular wedge pressure measurements showed comparable cardiac output (CO) and cardiac index (CI) between the two groups (P>0.05), while pulmonary capillary wedge pressure (PCWP) and pulmonary vascular resistance (PVR) showed significant differences (P<0.05). TTE assessment showed no significant differences in left ventricular (LV) size, right ventricular (RV) and right atrium (RA) measurements, or left ventricular ejection fraction (LVEF). However, the PH group exhibited significantly higher pulmonary artery systolic pressure (PASP) and tricuspid regurgitation velocity (TRV) (P<0.05), as shown in Table 2.

[0080] characteristic Total (n=51) TDL containing pH (n=31) TDL (non-PH) (n=20) P Age (years) 64.±8 63±8 65±8 0.566 Male, gender (%) 43(84.31) 16(80.00) 27(87.10) 0.496 BMI (kg / m2) 20.02±4.11 20.00±4.07 20.04±4.27 0.972 *mBP (mmHg) 95.24±14.15 95.19±14.60 95.32±13.77 0.977 θRR 20.10±3.70 20.61±3.84 19.30±3.39 0.219 Peripheral pulse oxygenation (%) 97.37±2.09 97.55±2.41 97.10±1.48 0.460 Course (y) 24.04±13.74 24.61±14.14 23.15±13.42 0.715 Comorbidities: COPD (%), pleurisy (%), CHD (%), hypertension (%) 9(17.65)11(21.57)6(11.76)13(25.49) 8(25.81)7(22.58)2(6.45)8(25.81) 1(5.00)4(20.00)4(20.00)5(25.00) 0.0570.8270.1430.949 NT-proBNP (pg / ml) 848.6（333.65,1498.9） 1054.9(527.9,2514.3) 441.85(207.1,1355.9) 0.036 Coughing up blood, minor to moderate, large bowel movements 41（80.39%）11(21.57)15(29.41)15(29.41) 7(22.58)8(25.81)9(29.03) 4(20.00)7(35.00)9(29.03) 0.8270.4820.941 §BAE, number of times 1.49±1±0.81 1.61±0.95 1.30±0.47 0.127 ▲Lung function FVC (predicted percentage) 1-month FEV1 (predicted value %) FEV1 / FVC DLCO / VARV (%) DLCO (mmol / min / kpa) RV / TL (predicted percentage) TLC (predicted value %) 1.65±0.6851.6±19.311.04±0.6238.62±21.2261.42±15.611.25±1.25147.02±75.895.39±2.6374.60±37.9179.30±33.153 1.47±0.5641.86±10.010.82±0.4129.63±10.2456.07±14.461.11±0.39131.67±55.574.14±2.1583.56±49.6569.91±18.64 1.99±0.8069.86±19.671.44±0.7755.48±26.6071.45±13.081.40±0.24164.29±94.826.80±2.4964.52±15.7589.86±43.29 0.0810.0040.0180.0290.0200.0940.3940.0320.3170.257

[0081] Table 1

[0082] *mBP = mean brachial artery blood pressure; θRR = respiratory rate; FVC = forced vital capacity; §BAE = bronchial artery embolization; abbreviations for pulmonary function parameters: FEV1 = forced expiratory volume in one second; RV = residual volume; DLCO = carbon monoxide pulmonary diffusion capacity, absolute value; DLCO / VA = carbon monoxide alveolar volume diffusion capacity per liter, absolute value; RV / TLC = residual volume / total vital capacity ratio;

[0083] Total (n=51) TDL containing pH (n=31) TDL (non-PH) (n=20) P #RHCCO (L / min) CI (L / min / m2) mPAP (mmHg) PVR (Wood) PCWP (mmHg) 5.33±1.353.27±3.2721（18，27）2.08（1.63,3.1）10.04±3.89 5.26±1.333.31±0.7826(22.0,29.0)2.59(1.8,4.2)11.10±4.08 5.43±1.413.21±0.7918(16.3,19.0)1.79(1.2,2.1)8.40±2.91 0.6680.665＜0.001＜0.0010.014 ❆Echocardiographic LV(cm) 4.25±0.47 4.24±0.44 4.27±0.53 0.855 RA(cm) 3.12±0.74 3.22±0.82 2.87±0.47 0.087 RV (cm) 2.95±0.77 3.06±0.90 2.74±0.32 0.096 LVEF (%) 64.52±8.98 63.57±10.37 66.43±5.00 0.337 PASP (mmHg) 39.52±22.35 44.85±24.96 27.67±6.30 0.008 TRV (m / s) 1.09±1.04 1.31±1.14 0.64±0.63 0.018

[0084] Table 2

[0085] #Abbreviations for RHC measurement parameters: CO = Cardiac output; CI = Cardiac index; mPAP = Mean pulmonary artery pressure; PVR = Pulmonary vascular resistance; PCWP = Pulmonary capillary wedge pressure; PASP = Pulmonary artery systolic pressure. ❆Abbreviations for TTE measurement parameters: LV = Left ventricular end-diastolic diameter; RV = Right ventricular end-diastolic diameter; RA = Right atrial anteroposterior diameter; LVEF = Left ventricular ejection fraction; TRV = Tricuspid regurgitation velocity.

[0086] CT imaging features and AI calculation results of TDL:

[0087] This study analyzed data from 969 lung segments in 51 patients. When comparing data from the two groups within lung segments with the same anatomical structure, no significant differences were found in the mean volume, lung tissue volume, and lung air volume of a single lung segment (P>0.05). However, the ratio of pulmonary ventilation to pulmonary ventilation time (Vventilation time) differed significantly between the two groups. TDLT There was a significant difference, specifically, being {86.510 (54.8, 128.7) ml vs 195.520 (159.3, 450.1) ml} (P < 0.01), as shown in Table 3:

[0088] characteristic Total (n=51) TDL containing pH (n=31) TDLNon-PH (n=20) P <![CDATA[V TDLT (ml)]]> 164.73（92.5，305.89） 195.52(159.3,450.1) 86.51(54.8,128.7) ＜0.01 DLL, digital 1.863±0.75 1.70±0.66 1.97±0.80 0.216 R-UL (%) 30(58.82) 19(61.29) 11(55.00) 0.656 R-ML (%) 12(23.53) 9(29.03) 3(15.00) 0.249 R-DL (%) 6(11.76) 3(9.68) 3(15.00) 0.565 L-UL (%) 32(62.75) 22(70.97) 10(50.00) 0.131 L-DL (%) 8(15.69) 5(16.13) 3(15.00) 0.914 Emphysema and / or bullae (%) 23(45.10) 17(54.84) 6(30.00) 0.082 Traction bronchodilation (%) 32(62.75) 21(67.74) 11(55.00) 0.358 pleurisy(%) 11(21.57) 7(22.58) 4(20.00) 0.827 Aba 2.63±1.30 2.74±1.53 2.45±0.83 0.438

[0089] Table 3

[0090] DLL = Destroyed lung lobe; R-UL = Right upper lobe; R-ML = Right middle lobe; R-DL = Right lower lobe; L-UL = Left upper lobe + sublingual lobe; L-DL = Left lower lobe; ABA = Bronchial artery abnormality.

[0091] See Figure 3 V TDLT There was a moderately statistically significant positive correlation with mean pulmonary artery pressure (mPAP) (r=0.552, P<0.001). In contrast, V TDLT No significant correlation was found between pulmonary artery systolic blood pressure (PASP) (r=-0.01, P=0.969) or pulmonary vascular resistance (PVR) (r=0.032, P=0.882). Similarly, the number of lobar injuries was borderline positively correlated with mPAP (r=0.224, P=0.084), but not significantly correlated with PASP (r=-0.021, P=0.914). Furthermore, a borderline positive correlation was also observed between DLL and PVR (r=0.166, P=0.244).

[0092] Please see Figure 4 ROC curve analysis was used to evaluate V. TDLT Diagnostic efficacy of PSAP and its combined diagnostic protocols. Results showed: V TDLT Pulmonary hypertension diagnosis: Area under the curve (AUC) was 0.894 (95% confidence interval: 0.76–1), sensitivity 90.32%, specificity 80%; cutoff value was 133.98 mL. PSAP pulmonary hypertension diagnosis: Area under the curve (AUC) was 0.761 (95% confidence interval: 0.59–0.93), sensitivity 55%, specificity 100%; cutoff value was 37.5 mmHg. Combined diagnostic protocol (V TDLT+PSAP): Area under the curve (AUC) was 0.917 (95% confidence interval: 0.802-1), predictive probability was 0.51, Youden index was 0.789; sensitivity was 90%, specificity was 88.9%.

[0093] In summary:

[0094] 1. Social benefits: Screening for pulmonary hypertension by combining CT imaging with artificial intelligence algorithms does not change the original screening method of right heart catheterization. It provides TDL patients with a non-invasive, early, and accurate PH screening tool, reduces the need for RHC, and alleviates patient suffering and medical burden.

[0095] 2. Economic benefits: Reduces the number of invasive examinations, saves medical resources, and improves diagnostic and treatment efficiency.

[0096] 3. Technical effect: V TDLT It was significantly correlated with mPAP (r=0.552, p<0.001), and the AUC of the combination with PASP reached 0.917, with a sensitivity of 90% and a specificity of 88.9%, which was significantly better than the single indicator.

[0097] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based chest CT imaging system for diagnosing pulmonary tuberculosis and pulmonary hypertension, characterized in that, include: The data acquisition module is used to acquire chest CT images of the patient in a supine position during full inspiration and export the CT images in DICOM format; The data processing module uses AI algorithms to automatically segment the entire lung and divide lung lobes using AI image analysis software, based on bronchial anatomy; it sets the HU range to identify damaged lung tissue and calculates V. TDLT ; The quantitative analysis module is used to output the volume of damage, the number and distribution of affected lung lobes; generate three-dimensional reconstructed images and mark the damaged areas; and provide correlation analysis charts between the volume of damage and mPAP and PASP. Risk prediction module, the risk prediction module will V TDLT Combined with clinical parameters, a PH risk prediction model was constructed; The report output module is used to generate structured reports, including text descriptions, charts, and risk warnings.

2. The artificial intelligence chest CT imaging system for diagnosing pulmonary tuberculosis-pulmonary hypertension according to claim 1, characterized in that, The acquisition of chest CT images of the patient in a supine position during the full inspiration phase specifically includes: Chest CT images of the patient in supine position during full inspiration were acquired using a Siemens 32-slice spiral CT scanner, with a slice thickness of 1 mm, and exported in DICOM format; then imported into AI image analysis software.

3. The artificial intelligence chest CT imaging system for diagnosing pulmonary tuberculosis-pulmonary hypertension according to claim 1, characterized in that: The automatic whole-lung segmentation and lobar segmentation are specifically as follows: A deep learning model based on U-Net or a similar structure is used for automatic whole-lung segmentation; the lungs are further segmented into lobes and segments based on bronchial anatomy; the HU range is defined to identify damaged lung tissue, and V is calculated. TDLT .

4. The artificial intelligence chest CT imaging system for diagnosing pulmonary tuberculosis-pulmonary hypertension according to claim 3, characterized in that, The automatic whole-lung segmentation is specifically as follows: First, identify the high curvature regions in the lung CT image, calculate the maximum principal curvature value (i.e., the maximum normal curvature), and visualize it using color based on its absolute value. The two largest surface regions are identified and correspond to the convex regions of the left and right lungs, respectively. An axis-aligned bounding box is then constructed for each identified surface region. Subsequently, by checking the bounding boxes to which each surface region belongs, all remaining surface regions are classified as left and right lung regions.

5. The artificial intelligence chest CT imaging system for diagnosing pulmonary tuberculosis-pulmonary hypertension according to claim 3, characterized in that, The lung lobe segmentation is specifically as follows: First, the seed point within the nodule is determined, and then the anatomical structure is modeled using the sub-volume centered on the seed point. Then, an AI algorithm is applied to generate a triangular mesh surface based on the scalar field data within the sub-volume; The Coendlink shape classification method based on the "visibility" detection process is adopted to determine the surface region of the nodules and identify their corresponding data points through specific seed points.

6. The artificial intelligence chest CT imaging system for diagnosing pulmonary tuberculosis-pulmonary hypertension according to claim 5, characterized in that, The process of determining the nodule surface region and identifying its corresponding data points using a specific seed point is as follows: Multiple rays are emitted from the seed point s, and the intersection points (p1, p2...p...) of the rays with the triangular mesh generated by the MCA are calculated. m ); Secondly, the first intersection point between the ray and the triangular mesh is defined as a "visible" point, and subsequent intersection points are marked as "invisible". Then, points that are "invisible" to the seed point are filtered out. The shape index of visible points is calculated using an AI algorithm, and points classified as convex prisms are removed. Discontinuous regions formed by all visible points are considered as components of the nodule surface.

7. The artificial intelligence chest CT imaging system for diagnosing pulmonary tuberculosis-pulmonary hypertension according to claim 3, characterized in that: The calculation of V TDLT Specifically, it involves automatically calculating the V value for each lung segment using an efficient grid simplification algorithm. TDLT Values ​​and air content.

8. A method for diagnosing pulmonary tuberculosis-pulmonary hypertension using artificial intelligence chest CT imaging according to any one of claims 1-7, characterized in that, Includes the following steps: S1. Acquire DICOM data of the patient's chest CT images through the data acquisition module; S2, The data processing module uses AI algorithms to automatically segment the whole lung and divide lung lobes; S3. The quantitative analysis module automatically identifies and quantifies the volume of destroyed lung tissue (V) based on the Hounsfield unit threshold. TDLT ); S4, Risk Prediction Module will V TDLT Combined with clinical parameters, a PH risk prediction model was constructed; S5. The report output module outputs a visual report, including the volume of damage, lung lobe distribution, pH risk level, and diagnostic recommendations.

Images