Lung cancer clinical T staging method and system based on data driving

Abstract

The invention provides a lung cancer clinical T-staging method and system based on data driving, which quantitatively calculates division thresholds of GGO and solid components of a lung tumor in a CT image, obtains a quantitative feature considering the GGO and the solid components at the same time, and creates an improved clinical T-staging system (cTm) by using the quantitative feature. The cTm aims at more accurately reflecting the influence of GGO, providing finer prognosis layering and pathological grade prediction, guiding a more personalized treatment strategy, and finally improving the survival rate and life quality of a patient.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a data-driven clinical T staging system for lung cancer. Background Art

[0002] Currently, for patients with clinical stage I non-small cell lung cancer (NSCLC), although surgery is the main treatment, the postoperative survival rate of patients is still not ideal. On CT images, lung cancer tumors actually contain two components: solid and GGO. The traditional cT8 staging method is mainly based on the size of the solid component of the tumor, and the presence of GGO is usually associated with lower invasiveness and better survival prognosis, but the existing staging system lacks sufficient consideration for this.

[0003] Recent studies have attempted to include GGO in the analysis through manual or semi-automatic methods. However, the main disadvantages of these methods are: (1) High subjectivity: There is no recognized threshold for demarcation of solid and GGO components of the tumor. The definition of tumor components depends on the experience of radiologists, resulting in poor consistency among different doctors. (2) Limited clinical value: Tumor pathological grade and prognosis are crucial to clinical treatment decisions. Existing methods do not simultaneously consider the prediction accuracy of tumor component quantitative characteristics on tumor pathological grade and prognosis.

[0004] The present invention designs a data-driven clinical T staging system for lung cancer, which provides more sophisticated prognostic stratification and pathological grade prediction, guides more personalized treatment strategies, and ultimately improves patients' survival rate and quality of life. Summary of the invention

[0005] The present invention provides a data-driven clinical T staging method for lung cancer, comprising:

[0006] Step 1, obtaining a chest CT image from a CT scanner;

[0007] Step 2: Use the deep learning model nnU-Net to perform tumor segmentation;

[0008] Step 3, determine the optimal threshold;

[0009] Step 4: Calculate the C-index of each feature under the optimal threshold combination, and select the feature with the highest C-index as the best feature;

[0010] Step 5, cT8 refinement: Use the best features to modify the existing clinical stage cT8.

[0011] Further, step 3 includes:

[0012] Test threshold combination (Tm i ,Ts j), wherein Tm represents the GGO threshold, Ts represents the realness threshold, i, j represent sequence numbers, i is a natural number from 1 to 7, j is a natural number from 1 to 11, there are 77 threshold combinations in total, the GGO component is the area between the selected GGO threshold and the realness threshold, and the realness component is the part above the realness threshold;

[0013] Each threshold combination (Tm i ,Ts j ) to extract 19 quantitative features;

[0014] The C-index of the LASSO-based pathological and prognostic prediction model was calculated using cross-validation under different threshold combinations;

[0015] The threshold with the highest C-index is taken as the optimal threshold.

[0016] Further, step 3 includes:

[0017] The Tm ranges from -800HU to -500HU, increasing by 50HU each time;

[0018] The Ts ranges from -500HU to 0HU, increasing by 50HU each time.

[0019] Furthermore, the 19 quantitative features include:

[0020] Maximum total size (MTSA), maximum solid size (MSSA), proportion of solid size to total size (MSSA%), maximum GGO size (MGSA), maximum total size by volume (MTSV), maximum solid size (MSSV), maximum GGO size (MGSV), total volume (TV), solid component volume (SV), GGO component volume (GV), proportion of solid component volume to total volume (SV%), proportion of GGO component volume to total volume (GV%), total mass (TM), solid mass (SM), GGO mass (GM), proportion of solid component mass (SM%), proportion of GGO component mass (GM%), mean HU (MHU), entropy of HU (EHU).

[0021] Further, step 4 includes:

[0022] The quantitative features of each threshold combination were used in combination with the machine learning method LASSO regression to predict the pathological grade and prognosis. The C-index was used to quantitatively evaluate the prediction effect. After 5 cross-validations, the optimal threshold combination was determined based on the threshold combination with the highest C-index mean.

[0023] Further, step 4 includes:

[0024] The C-index of each feature under the optimal threshold combination was calculated, and the feature with the highest C-index was selected as the optimal feature for predicting pathological grade and prognosis. After incorporating other clinical factors, the correlation between the optimal feature and prognosis was obtained by Cox regression. The relationship between the optimal feature and pathological grade was analyzed by logistic regression. The correlation between the optimal feature and the size of pathological invasion was verified by linear regression and Pearson correlation coefficient.

[0025] Further, based on the correlation, the best features were used to modify the clinical stage cT8:

[0026] The patients are downstaged or upstaged according to the characteristic values ​​to obtain a modified clinical T stage, i.e., cTm. The clinical T stage of patients with characteristic values ​​higher than a determined threshold T_up is adjusted upward to an adjacent stage, and the clinical T stage of patients with characteristic values ​​lower than a determined threshold Tdown is adjusted downward to an adjacent stage. The thresholds T_up and T_down are determined by a traversal search method, and the optimal threshold combination T_up and T_down is defined as the one that enables cTm to obtain the best performance in pathological grade and prognosis, with the highest C-index.

[0027] The present invention also provides a data-driven clinical T staging system for lung cancer. The optimal thresholds for distinguishing GGO and solid components on CT images are -650HU and -250HU, respectively. The SM% calculated based on the thresholds is the optimal feature for predicting pathological grading and prognosis. By downgrading the cT8 stage for patients with SM% below 45% and upgrading the cT8 stage for patients with SM% above 75%, the C-index of pathological and prognostic prediction can be significantly improved, p<0.001.

[0028] The technical effects of the present invention are:

[0029] (1) Determination of optimized component division thresholds: By testing 77 different combinations of GGO and solid component division thresholds, the present invention determined -650HU and -250HU as the optimal thresholds, which, compared with traditional fixed thresholds, can better reflect the true composition of the tumor, thereby improving the accuracy of pathological grading and prognosis prediction.

[0030] (2) Improved clinical T staging system cTm: The present invention proposes cTm by adjusting the clinical T staging (cT8), which reclassifies patients based on the solid component percentage (SM%), downgrading the staging for patients with SM% below 45% and upgrading the staging for patients with SM% above 75%. Compared with cT8, which only relies on tumor size, this data-driven T staging adjustment can more accurately reflect tumor biological behavior and prognosis, thereby guiding more reasonable treatment decisions.

[0031] (3) Significantly improved predictive performance: In multiple validation cohorts, cTm showed significantly better performance than cT8 in pathological grading and prognostic prediction, as reflected by higher C-index values, which was due to the fact that the present invention more comprehensively considered the potential impact of GGO components on prognosis.

[0032] (4) Clinical decision optimization: The cTm provided by the present invention helps to more accurately stratify patient risks. For patients with clinical stage I lung adenocarcinoma, it can guide more personalized treatment strategies. For example, patients with up-staged disease may require more aggressive treatment plans, while patients with down-staged disease may consider more conservative management. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Shown is a flow chart of image contrast adjustment according to the present invention. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] See also Figure 1 The present invention provides a data-driven quantitative analysis method to improve the clinical T staging (cT8) of clinical stage I lung adenocarcinoma. The present invention also provides a data-driven clinical T staging system for lung cancer, which aims to optimize the prediction ability of pathological grading and recurrence-free survival (RFS) by extracting quantitative features including GGO and solid components.

[0036] First, chest CT images were obtained from a CT scanner, and the longest diameter of the tumor and the longest diameter of the solid component were measured using the commonly used clinical window settings (window level -500 Hounsfield units, HU; window width 1500 HU). Then, the deep learning model nnU-Net was used to perform tumor segmentation.

[0037] Since there is no universally accepted threshold to distinguish GGO from solid components, 77 threshold combinations were tested, including 7 different GGO thresholds (from -800HU to -500HU, with an increase of 50HU each time) and 11 different solid thresholds (from -500HU to 0HU, with an increase of 50HU each time). We defined the GGO component as the area between the selected GGO threshold and the solid threshold, while the solid component refers to the part above the solid threshold.

[0038] For each threshold combination, a series of quantitative features were calculated, including maximum total size (MTSA), maximum solid size (MSSA), proportion of solid size to total size (MSSA%), maximum GGO size (MGSA), maximum total size by volume (MTSV), maximum solid size (MSSV), maximum GGO size (MGSV), total volume (TV), solid component volume (SV), GGO component volume (GV), proportion of solid component volume to total volume (SV%), proportion of GGO component volume to total volume (GV%), total mass (TM), solid mass (SM), GGO mass (GM), proportion of solid component mass (SM%), proportion of GGO component mass (GM%), mean HU (MHU), and entropy of HU (EHU). The definitions and formulas of each quantitative feature are as follows:

[0039] 1. Maximum total size (MTSA): defined as the maximum diameter of the entire tumor (including GGO and solid components) on axial CT images.

[0040] 2. Maximum solid size (MSSA): defined as the maximum diameter of the solid component of the tumor on axial CT images.

[0041] 3. Ratio of solid size to total size (MSSA%): defined as the ratio of the maximum solid size (MSSA) to the maximum total size (MTSA), formula: MSSA% = (MSSA / MTSA) × 100%.

[0042] 4. Maximum GGO size (MGSA): defined as the maximum diameter of the GGO component of the tumor on axial CT images.

[0043] 5. Maximum total size by volume (MTSV): defined as the total size of the entire tumor (including GGO)

[0044] Formula: Calculated based on 3D reconstruction.

[0045] 6. Maximum solid size (MSSV): defined as the three-dimensional volume of the solid component of the tumor. Formula:

[0046] Calculated based on 3D reconstruction.

[0047] 7. Maximum GGO size (MGSV): defined as the three-dimensional volume of the GGO component of the tumor. Formula: calculated based on 3D reconstruction.

[0048] 8. Total volume (TV): defined as the total volume of the tumor, including solid and GGO components.

[0049] 9. Solid component volume (SV): defined as the volume of the solid component of the tumor.

[0050] 10. GGO component volume (GV): defined as the volume of the GGO component of the tumor.

[0051] 11. Solid component volume to total volume ratio (SV%): defined as the ratio of solid component volume (SV) to total volume (TV). Formula: SV% = (SV / TV) × 100%

[0052] 12. The ratio of GGO volume to total volume (GV%): defined as the ratio of GGO volume (GV) to total volume (TV). Formula: GV% = (GV / TV) × 100%

[0053] 13. Total mass (TM): defined as the total mass of the tumor, including the mass of the solid and GGO components. Formula: TM = density × TV (converted to density value in Hounsfield units (HU)).

[0054] 14. Solid mass (SM): defined as the mass of the solid component of the tumor. Formula: SM = density × SV

[0055] 15. GGO mass (GM): defined as the mass of the GGO component of the tumor. Formula: GM = density × GV.

[0056] 16. Solid component mass ratio (SM%): defined as the ratio of the solid component mass (SM) to the total mass (TM). Formula: SM% = (SM / TM) × 100%.

[0057] 17. GGO mass percentage (GM%): defined as the ratio of the mass of GGO (GM) to the total mass (TM). Formula: GM% = (GM / TM) × 100%

[0058] 18. Mean HU (MHU): represents the average CT value (Hounsfield Unit) of the entire lesion area, which is the average of the HU values ​​of all voxels.

[0059] 19. Entropy of HU (EHU): used to measure the uniformity or uncertainty of the distribution of HU values ​​in CT images. Formula: EHU = -∑ipi·log2(pi) EHU = -∑ipi·log2(pi) Where pi is the probability of a voxel appearing in the i-th HU value interval, and log2(pi) is the logarithm of the probability pi to the base 2.

[0060] The quantitative features under each threshold combination were combined with the machine learning method LASSO regression to predict pathological grade and prognosis. The C-index was used to quantitatively evaluate the prediction effect. After 5 cross-validations, the best threshold combination was determined by the threshold combination with the highest C-index mean.

[0061] The C-index of each feature under the optimal threshold combination was calculated, and the feature with the highest C-index was selected as the optimal feature for predicting pathological grade and prognosis. After incorporating other clinical factors, the correlation between the optimal feature and prognosis was obtained by Cox regression. Logistic regression was used to analyze the relationship between the optimal feature and pathological grade. The correlation between the optimal feature and the size of pathological invasion was verified by linear regression and Pearson correlation coefficient.

[0062] Based on the correlation found in the above analysis, the optimal feature is used to modify the existing clinical stage cT8: the patient is downstaged or upstaged according to the feature value to obtain the modified clinical T stage (i.e., cTm). Specifically, the clinical T stage of patients with feature values ​​higher than the determined threshold (T_up) is adjusted upward to the adjacent stage, and the clinical T stage of patients with feature values ​​lower than the determined threshold (T_down) is adjusted downward to the adjacent stage. The thresholds T_up and T_down are determined by the traversal search method, and the optimal threshold combination T_up and T_down is defined as the one that enables cTm to obtain the best performance of pathological grade and prognosis (highest C-index).

[0063] The improved clinical T staging system (cTm) obtained based on the above data-driven method: the optimal thresholds for distinguishing GGO and solid components on CT images are -650HU and -250HU, respectively. The SM% calculated based on the threshold is the optimal feature for predicting pathological grade and prognosis. By downgrading the cT8 stage for patients with SM% below 45% and upgrading the cT8 stage for patients with SM% above 75%, the C-index for pathological and prognostic prediction can be significantly improved (p<0.001).

[0064] Compared with the prior art, the present invention is different in the following key aspects:

[0065] (1) Quantification of GGO and solid components: The present invention uses nnUnet model-assisted automatic segmentation and extensive threshold testing to determine -650HU and -250HU as the optimal thresholds for GGO and solid components. This is different from the method of using fixed thresholds based on subjective experience (such as -300HU to 0HU) in previous studies, providing more accurate quantification and improving the accuracy of pathology prediction and RFS prediction.

[0066] (2) Improvement of clinical T staging of lung cancer: Based on the quantified GGO and solid components, the present invention proposes cTm, which reclassifies patients according to the proportion of solid components to total components (SM%) by adjusting the existing clinical T staging cT8. cTm downgrades patients with SM% less than 45% and upgrades patients with SM% greater than 75%. Compared with the existing clinical T staging cT8, it significantly improves the C-index and optimizes the pathological grading and prognosis prediction capabilities.

[0067] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0068] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is limited by the appended claims and their equivalents.

Claims

1. A data-driven clinical T staging method for lung cancer, characterized in that: include: Step 1, obtaining a chest CT image from a CT scanner; Step 2: Use the deep learning model nnU-Net to perform tumor segmentation; Step 3, determine the optimal threshold; Step 4: Calculate the C-index of each feature under the optimal threshold combination, and select the feature with the highest C-index as the best feature; Step 5, cT8 refinement: Use the best features to modify the existing clinical stage cT8.

2. A data-driven clinical T staging method for lung cancer according to claim 1, characterized in that: Step 3 includes: Test threshold combination (Tm i ,Ts j ), wherein Tm represents the GGO threshold, Ts represents the realness threshold, i, j represent sequence numbers, i is a natural number from 1 to 7, j is a natural number from 1 to 11, there are 77 threshold combinations in total, the GGO component is the area between the selected GGO threshold and the realness threshold, and the realness component is the part above the realness threshold; Each threshold combination (Tm i ,Ts j ) to extract 19 quantitative features; The C-index of the LASSO-based pathological and prognostic prediction model was calculated using cross-validation under different threshold combinations; The threshold with the highest C-index is taken as the optimal threshold.

3. A data-driven clinical T staging method for lung cancer according to claim 2, characterized in that: Step 3 includes: The Tm ranges from -800HU to -500HU, increasing by 50HU each time; The Ts ranges from -500HU to 0HU, increasing by 50HU each time.

4. A data-driven clinical T staging method for lung cancer according to claim 3, characterized in that: The 19 quantitative characteristics include: Maximum total size (MTSA), maximum solid size (MSSA), proportion of solid size to total size (MSSA%), maximum GGO size (MGSA), maximum total size by volume (MTSV), maximum solid size (MSSV), maximum GGO size (MGSV), total volume (TV), solid component volume (SV), GGO component volume (GV), proportion of solid component volume to total volume (SV%), proportion of GGO component volume to total volume (GV%), total mass (TM), solid mass (SM), GGO mass (GM), proportion of solid component mass (SM%), proportion of GGO component mass (GM%), mean HU (MHU), entropy of HU (EHU).

5. A data-driven clinical T staging method for lung cancer according to claim 4, characterized in that: Step 4 includes: The quantitative features of each threshold combination were used in combination with the machine learning method LASSO regression to predict the pathological grade and prognosis. The C-index was used to quantitatively evaluate the prediction effect. After 5 cross-validations, the optimal threshold combination was determined based on the threshold combination with the highest C-index mean.

6. A data-driven clinical T staging system for lung cancer according to claim 5, characterized in that: Step 4 includes: The C-index of each feature under the optimal threshold combination was calculated, and the feature with the highest C-index was selected as the optimal feature for predicting pathological grade and prognosis. After incorporating other clinical factors, the correlation between the optimal feature and prognosis was obtained by Cox regression. The relationship between the optimal feature and pathological grade was analyzed by logistic regression. The correlation between the optimal feature and the size of pathological invasion was verified by linear regression and Pearson correlation coefficient.

7. A data-driven clinical T staging system for lung cancer according to claim 6, characterized in that: Based on the described correlations, the best features were used to modify the clinical stage cT8: The patients are downstaged or upstaged according to the characteristic values ​​to obtain a modified clinical T stage, i.e., cTm. The clinical T stage of patients with characteristic values ​​higher than a determined threshold T_up is adjusted upward to an adjacent stage, and the clinical T stage of patients with characteristic values ​​lower than a determined threshold T_down is adjusted downward to an adjacent stage. The thresholds T_up and T_down are determined by a traversal search method, and the optimal threshold combination T_up and T_down is defined as the one that enables cTm to obtain the best performance in pathological grade and prognosis, with the highest C-index.

8. A system for clinical T staging of lung cancer based on data-driven method according to any one of claims 1 to 7, characterized in that: The optimal thresholds for distinguishing GGO and solid components on CT images are -650HU and -250HU, respectively. The SM% calculated based on the threshold is the optimal feature for predicting pathological grade and prognosis. By downgrading the cT8 stage for patients with SM% below 45% and upgrading the cT8 stage for patients with SM% above 75%, the C-index for pathological and prognostic prediction can be significantly improved, p<0.001.

Images