A prognostic marker for neoadjuvant immunotherapy and chemotherapy of lung cancer and application thereof

Abstract

The present application provides a prognostic marker for neoadjuvant immunotherapy and chemotherapy of lung cancer, which is the change rate of tumor volume before and after early treatment of lung cancer. In the early treatment of lung cancer, by combining the method of artificial intelligence, the volume change of tumor lesion is used as the feature of treatment effect, and it is observed from the Kaplan-Meier curve that the change amount of tumor volume has a more obvious difference between the two groups of patients.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of medical detection, and particularly relates to a prognostic marker for neoadjuvant immunotherapy and chemotherapy of lung cancer and application thereof. BACKGROUND

[0002] Lung cancer is one of the most common malignant tumors and is the leading cause of cancer-related deaths. Most lung cancers belong to non-small cell lung cancer (NSCLC). Chemotherapy and immunotherapy as independent therapies play an important role in the treatment of NSCLC, and are recommended to be used in different circumstances according to the ASCO (American Society of Clinical Oncology), ESMO (European Society for Medical Oncology) and NCCN (National Comprehensive Cancer Network) guidelines. The prognosis (five-year survival rate) of patients with advanced NSCLC has been improved from about 15% to 42.9%.

[0003] In order to better monitor the disease progression of patients, tumor response is designed as an expected endpoint of early clinical trials. In 2000, Therasse et al. proposed the Response Evaluation Criteria in Solid Tumors (RECIST) guideline, which uses the measurement of tumor diameter to present objective response (OR) and disease progression. Later, more studies claimed that the measurement of CT RECIST response by tumor diameter change reflected the objective response rate (ORR) and overall survival (OS) of NSCLC patients. However, there is a significant inconsistency between CT RECIST response and other prognostic indicators (such as pathological response), and the inconsistency rate is about 41% to 45%. It is also reported that in two cases, CT RECIST response has no significant correlation with overall survival (P value of CT RECIST response is 0.23 at 50% threshold, and P value of CT RECIST response is 0.98 at 70% threshold).

[0004] Travis et al. proposed pathological response. The percentage of viable tumor tissue, necrosis and stroma was used to indirectly assess the overall survival of all systemic treatments, especially in resectable NSCLC patients receiving neoadjuvant therapy. They defined major pathological response (MPR) and pathological complete response (pCR) according to the percentage of surviving tumor tissue. According to the study, the reason for this inconsistency is that the appearance of necrosis and stroma is similar to the viable tumor observed by CT. Therefore, the reduction of surviving tumor cannot be observed on radiological images, which confuses radiological assessment.

[0005] Several studies have shown that more accurate radiological indicators, such as volume response measurement, can more accurately reflect pathological response and OS. However, due to the limitations of measurement techniques, no new indicators have been proposed so far. Therefore, how to find a new CT RECIST response indicator in the early treatment stage (the first four treatment stages) is still a technical problem to be solved by those skilled in the art. SUMMARY

[0006] To solve the various problems existing in the prior art, therefore, we tried to find a new CT RECIST response indicator in the early treatment stage (first four treatment stages) to accurately indicate prognosis using an AI-based method. In our experiment, we first investigated irregularities in 916 tumor lesions during treatment. Then, we also compared changes in tumor diameter and volume in the early treatment stage. In addition, to reveal the CT RECIST response (corresponding to ΔD, ΔV) measured by changes in tumor diameter / volume in the early treatment stage and the pathological response. We also investigated their relationship with overall survival, as shown in Figure 1 (a).

[0007] Without being bound by any theory, the present inventors surprisingly found that the rate of change in tumor volume before and after early treatment of lung cancer has a strong correlation with the CT RECIST response indicator and overall generation rate, and thus completed the present application.

[0008] The present application provides a marker for neoadjuvant immunotherapy and chemotherapy of lung cancer, which is the rate of change in tumor volume before and after early treatment of lung cancer.

[0009] Another aspect of the present application also relates to the use of the rate of change in tumor volume before and after early treatment of lung cancer in the preparation of a diagnostic device for the prognosis of neoadjuvant immunotherapy and chemotherapy of lung cancer.

[0010] Another aspect of the present application relates to a diagnostic device, which comprises a CT, characterized in that the device has a module for calculating the rate of change in tumor volume before and after early treatment of lung cancer based on the CT diagnosis result.

[0011] In a preferred embodiment of the present application, the diagnostic device comprises the following modules:

[0012] (1) a lesion segmentation module, which, for lung cancer CT images in dicom format, uses a pre-trained NNU-net deep model framework based on lung cancer CT image data in the Medical Segmentation Decathlon (MSD) dataset, and uses the pre-trained model for CT image data segmentation;

[0013] (2) a lesion optimization module, which, for the lesions segmented based on NNU-net, uses a Hounsfield Units (HU) threshold method to remove pixel points with a Hounsfield Units value less than -100, and obtains the main lesion with the largest connected domain;

[0014] (3) a lesion volume calculation module, which uses open-source pyradiomics to obtain the diameter and volume of the tumor, wherein the tumor diameter is defined as the maximum pairwise Euclidean distance between the grid vertices on the tumor surface, and the tumor volume is calculated by the Marching Cubes algorithm.

[0015] In a preferred embodiment of the present application, the calculation of the tumor volume comprises the following steps: based on a 2x2 cube, N f tetrahedrons are found for each triangular mesh, and the volume V i of each tetrahedron i is calculated by the following formula

[0016]

[0017] where a i ,b i and c i represent the lengths of the three edges of the triangular mesh; the final tumor volume value is obtained by summing the volumes of all tetrahedrons using the summation formula.

[0018] In a preferred embodiment of the present application, since there are a portion of JPG format lung cancer CT images in the data set, the pre-trained model provided by NNU-net is directly used for lesion segmentation when performing lesion segmentation on this batch of lung cancer CT image data, and the segmentation result is poor. In order to solve the Domain Adaptation problem existing between the training set and the test set, the MSD training data is subjected to min-max normalization operation to normalize its data range to 0-255. Then the pre-trained NNU-net model is fine-tuned using the normalized data set, and finally the fine-tuned model is applied to the JPG format lung cancer data set to obtain the corresponding lesion segmentation result.

[0019] Advantages

[0020] Compared with the prior art, the present application has at least the following beneficial effects:

[0021] The present application, in the early stage of treatment, takes the volume change of tumor lesions as a feature of treatment efficacy by combining artificial intelligence: through experimental observation, it is found that the change of the maximum diameter of the tumor in the early stage of treatment cannot accurately reflect the final pathological response. The change in the volume of the tumor relative to the change in the maximum diameter has a strong correlation with the residual tumor volume of the pathological section. And in the process of survival analysis, we also observed from the Kaplan-Meier curve that the change in the volume of the tumor has a more obvious difference between the two groups of patients. Although it is difficult for clinicians to obtain the volume variable of the tumor through visual observation, the feature value can be more accurately obtained through artificial intelligence means. Therefore, we propose that the volume change of the tumor lesion is taken as one of the features of treatment efficacy by combining artificial intelligence. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 Figure 1 (a) is the analysis process, and the two kinds of relationships corresponding to ① and ② in the figure are observed. Figure 1 (b) is a patient inclusion flowchart, which shows the selection criteria and the number of patients in each part.

[0023] Figure 2 Figure 2 (a) is the pre-treatment / post-treatment CT image and pathological image of 3 patients. Each row represents a patient. For the patient in the first row, the change in the diameter on the axial slice is -15.48%, and the pathological response is SD. In the second and third rows, the change in the diameter along the axial slice is -15.71% and -10.95%, and the pathological evaluation results are both MPR. Figure 2 (b) is the CT and pathological image of a patient receiving immunotherapy treatment. The CT images before and after immunotherapy show a large ΔV (-61.97%), but a small ΔD (-2.64%). Before immunotherapy, the tumor diameter and volume were 5.68 cm and 38.29 cm 3 respectively. However, after immunotherapy, the tumor diameter is still 5.53 cm, but the volume has been reduced to 14.56 cm 3 . Postoperative pathology confirmed that the residual viable tumor was 6%, showing that the patient was MPR. Figure 2 (c) is the pipeline segmentation and diameter / volume calculation of the tumor lesion. A pre-trained nnU-Net model is used to infer the segmentation results of our CT images. Then the pixel points with HU value lower than -100 are removed from the segmentation results. After that, we should get the lesion with the largest connected domain, ignoring small and isolated tumor lesions. Finally, two professional radiologists are invited to check the segmentation results and refine the segmentation results.

[0024] Figure 3 ​​: Scatter plot between D and V of 916 tumor lesions. In the plot, the dots represent each patient, and the green curve shows the relationship between the diameter and volume of a standard sphere, in which the SCC between D and V is 1.0, and 0.9222 in tumor lesions.

[0025] Figure 4 : Waterfall plot of ΔD and ΔV of 136 patients.

[0026] Figure 5 : ΔD, ΔV and pathological response of 35 patients. In the left sub-plot, patients who only received chemotherapy, and in the right sub-plot, patients who received neoadjuvant immunotherapy combined with chemotherapy. The numbers listed on the bar chart represent the values of ΔD, ΔV and the percentage of residual surviving tumor. The colors represent the response defined by CT RECIST and pathological assessment. The best cutoff is the threshold value adjusted according to the volume.

[0027] Figure 6 : Survival analysis curve of 47 patients with advanced stage. DETAILED DESCRIPTION

[0028] The technical solutions in the embodiments of the present application will be clearly and completely described below in combination with specific embodiments of the present application, so as to help better understand the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments.

[0029] Unless otherwise specified, the reagents involved in the embodiments of the present application are commercially available products.

[0030] Example 1

[0031] 1. Dataset collection

[0032] In our study, 916 tumor lesions of 185 patients were collected. From these patients, we retrospectively retrieved and collected information of 136 patients who underwent CT images before first treatment and CT images obtained for the first time after early treatment from December 2015 to January 2022 in the following medical centers: General Hospital of People's Liberation Army, Beijing Shijitan Hospital Affiliated to Capital Medical University, Qingdao University Affiliated Hospital. The detailed information of these patients is included in Supplementary Table 1, and the characteristics of these patients are listed in Table 4, including age and gender. There are 43 females and 93 males, with an average age of 59.97 years old. Among the total of 136 patients, 89 patients received neoadjuvant treatment, and another 47 patients underwent adjuvant chemotherapy. In addition, we followed up the survival rate of 47 patients who received adjuvant chemotherapy. The follow-up period started from the day of immunotherapy or chemotherapy, and the overall survival was defined as the length of time (in days) from the day of immunotherapy and chemotherapy to the last day of follow-up or death.

[0033] Patient characteristics

[0034] Figure 1 (b) A flowchart of patient inclusion is given. We collected 916 tumor lesions from 185 patients. Among them, 136 patients (272 tumor lesions) had two CT scans, one before the first treatment and one after the first treatment. Then, from the selected 136 patients, we further picked 35 patients with pathology images and 47 advanced patients with complete follow-up records. Table 1 and Table 2 list their characteristics. Table 1 provides information of the 35 pathology image patients, including age, gender, histology type, treatment type, and pathological response. Among the 35 patients, 17 are lung squamous cell carcinoma (LUCS) and 14 are lung adenocarcinoma (LUAD). The remaining 4 patients are adenoid squamous carcinoma for 3 and fibrosarcoma for 1. Among the 35 patients, 20 patients received immunotherapy combined with chemotherapy, and the other 15 patients received chemotherapy alone. Regarding pathological response, we have 7 patients labeled as pCRs and 10 patients labeled as MPRs. We classify these 17 patients into the MPR category, indicating good pathological response, while the remaining 18 patients belong to the non-MPR, indicating poor pathological response. Table 2 provides information of the 47 advanced patients, with an average OS of 905.62 days and an average age of 62.83 years.

[0035] Table 1 Pathology image characteristics of 35 patients (including age, gender, histology type, treatment type, and pathological response information)

[0036]

[0037] Table 2 Characteristics of 47 advanced OS patients (including age, gender, and OS information)

[0038]

[0039] 2Radiological assessment

[0040] For each patient, the CT image taken before the first treatment was selected as his / her pre-treatment image, while the first CT image obtained after the early treatment stage was selected as the post-treatment image. In addition, all these CT images were obtained before surgery. We calculated the tumor volume and diameter and their changes from all these CT images. According to the standard of RECIST, complete remission (CR) means the disappearance of all target lesions, while partial remission (PR) occurs when the diameter of target lesions is reduced by at least 30%. Progressive disease (PD) refers to an increase in the diameter of all target lesions by more than 20%, while stable disease (SD) is between PR and PD.

[0041] To extract the diameter and tumor volume from CT images, we first used nnU-Net, a deep learning-based segmentation method to identify the approximate location and shape of tumor lesions. In this step, we downloaded a pre-trained model based on a public dataset, on top of which we could infer the segmentation results of our CT images. Considering that pixels with Hounsfield Units (HU) lower than -100 are mainly air and fat, we removed pixels with HU lower than the threshold of -100. Since some patients can have multiple small tumor lesions, we chose to focus on the largest connected component, while ignoring smaller and isolated tumor lesions. We then invited two professional radiologists to check the segmentation results and refine the segmentation results. The segmentation process of tumor lesions is shown in Fig. Figure 2 (c).

[0042] After that, we selected an open-source Python package pyradiomics to derive 3D shape features, including 3D diameter and mesh volume. A resampling operation was adopted to ensure that the isotropic voxels have equal distances between neighboring voxels in all directions. The 3D diameter was defined as the maximum pairwise Euclidean distance between mesh vertices on the tumor surface, called the Feret diameter. To calculate the volume, the 2x2 cubic marching cubes algorithm was used to find N f Triangles were used to construct the mesh. Each tetrahedron was constructed from the corresponding triangles in the mesh, with volume V i which can be calculated as Vi = a i x b i x c i / 6, where a i , b i and c i denote the lengths of the three edges forming the triangle. Then all Vi were summed up to calculate the mesh volume V. The changes in tumor diameter and volume were defined as ΔD = (D post - D pre ) / D pre , ΔV = (V pre - V post ) / V pre , where pre and post denote pre-treatment and post-treatment, respectively.

[0043] 3 Pathological evaluation

[0044] Surgical specimens were fixed in 10% formalin, embedded in paraffin, sectioned with a microtome, and stained with hematoxylin and eosin (H&E). According to the IASLC multidisciplinary recommendations, the area where the primary tumor was located before treatment was identified as the tumor bed. All pathological sections were scanned by Pannoramic MIDI digital slide scanner (3D HISTECH, Hungary). CaseViewers software was used to review and determine the residual viable tumor in the primary tumor. Two experienced senior pathologists reviewed and assessed the objective pathological response to treatment without knowing the treatment strategy and prognosis. If there was a disagreement in the evaluation, they reevaluated the answer in a consensus manner. The tumor bed included viable tumor, necrosis, and stroma, which included fibrosis and inflammation, and the three components added up to 100%. The percentage of viable tumor tissue was recorded for each tumor slide, and the final pathological response depended on the average result of all slides for each patient. According to the IASLC multidisciplinary recommendations, MPR was defined as 10% or less residual viability in the primary tumor, regardless of whether there were residual cancer cells in the lymph nodes, and pCR meant no viable tumor cells in all slides. Histological subtypes were determined according to the 2015 World Health Organization Classification of Tumors.

[0045] 4 Statistical analysis

[0046] Statistical data analysis was performed using the software RStudio 3.6.2. Spearman’s correlation coefficient (SCC), Pearson’s correlation coefficient (PCC), P-value, and R-squared were used to evaluate the relationship between early tumor / volume changes and OS. Log-rank test and Kaplan-Meier analysis were used to estimate the survival curve of recurrence-free survival probability. A P-value less than 0.05 meant that the difference was statistically significant.

[0047] 5 Case presentation: RECIST vs. pathological evaluation inconsistency

[0048] In Figure 2 , we first show the pre-treatment / post-treatment CT images and their pathological images of three patients who received neoadjuvant immunotherapy combined with chemotherapy, as shown in Figure 2 (a), where each row represents a patient. Since it is difficult for clinicians to calculate the diameter from 3D shapes, they usually calculate the diameter from the axial slice with the largest cross-sectional area of the tumor. For the patient in the first row of Figure 2 (a), we can see that the change in diameter on the axial slice is -15.48%. According to RECIST, the patient would be classified as SD. However, according to the pathological evaluation, the patient belongs to pCR, which means that the tumor lesion has completely disappeared. Similarly, for the patient in the second row of Figure 2(a) The second and third rows of patients, the changes in diameter on axial slices were -15.71% and -10.95%, respectively, and thus were considered as SD according to RECIST, while the pathological assessment results were both MPR, which means that the tumor lesions had mostly disappeared. These cases suggest that AD is not accurate enough. Therefore, we explored the pathological response of AD / AV at the early treatment stage and found a new CT RECIST response indicator in the following experiment, which accurately reflects the prognosis in the early treatment.

[0049] 6Relationship between D and V in tumor lesions

[0050] Given the inconsistency between CT RECIST and pathological assessment, we first explored the relationship between D and V during treatment, as shown in Figure 3 We also calculated the Spearman correlation coefficient (SCC) between them. Then we observed AD and AV of each patient at the early treatment stage of tumor lesions for 136 patients, as shown in Figure 4 The characteristics of the 136 patients are shown in Table 3.

[0051] Table 3 Characteristics of 136 patients, including information on age, gender, and treatment type

[0052]

[0053] In Figure 3 , each point represents a tumor lesion. The green curve shows the relationship between the diameter and volume of a standard sphere. As can be seen from Figure 3 , although D and V in tumor lesions are similar to spheres, the SCC of D and V in tumor lesions (0.9222) is lower than that in standard spheres (SCC = 1.0), indicating that the correlation between the two variables in tumor lesions is weaker. This finding suggests that the shape of the tumor lesion may be irregular, deviating from the spherical shape of the standard sphere.

[0054] Then to analyze the relationship between AD and AV at the early treatment stage in detail, we used the waterfall plot in Figure 4 to visualize AD and AV. In this figure, the red bars represent AD and the blue bars represent AV. In Figure 4In the middle, according to the thresholds of 20% and -30% mentioned in ECIST, it can be divided into 5 parts. In the first part, both AD and AV are greater than -30% but less than 20%. According to RECIST, these patients are classified as SD regardless of the index AD or AV. In the second part, the reduction of diameter in these tumor lesions has not reached 30%, while the reduction of tumor volume has exceeded 30%, which are inconsistent. The same situation can also be observed in the third part, where the increase of diameter has not reached 20%, while the increase of tumor volume has exceeded 20%. In the fourth part, AD and AV are consistent with each other. Unlike the fourth part, AD and AV are completely opposite in the last part. Specifically, in the last part, the diameter can only increase slightly, while the volume of the tumor is at least reduced by 30%, or vice versa.

[0055] By analyzing the data of Figure 3 and Figure 4 , we observed 916 tumor lesions irregularities in the early treatment stage and the inconsistency between AD and AV. In the subsequent experiments, we further investigated the AD and AV in the early treatment stage and pathological response. In addition, we also investigated their relationship with overall survival.

[0056] 7Relationship between AD, AV and pathological response

[0057] After observing the irregularities of all tumor lesions and the inconsistency between AD and AV in the early treatment stage, we explored the AD / AV in the early treatment stage and detailed pathological response. In this experiment, 35 patients with corresponding pathological images were selected, and their characteristics are listed in Table 1. We measured AD and AV and the percentage of residual viable tumor observed on the pathological image. Then, we also investigated the correlation between AD, AV and the percentage of residual viable tumor on the pathological image in the early treatment stage, and the results are shown in Figure 5 .

[0058] First, we created a visualization of the three variables, i.e., AD, AV and the percentage of residual viable tumor. The left subgraph refers to patients who only received chemotherapy, and the right subgraph refers to patients who received neoadjuvant immunotherapy combined with chemotherapy. In the first row of the two subgraphs, the numbers listed on the bar chart are the changes in the percentage of residual viable tumor on the pathological image. Red represents MPR patients whose residual viable tumor percentage on the pathological image has decreased by more than 90%, while blue represents non-MPR patients. Note that MPR patients refer to patients with positive treatment response, and non-MPR patients refer to patients with negative treatment response. Similarly, the numbers listed on the bar charts in the second and third rows refer to AD and AV, respectively. The colors in the second and third rows represent AD and AV, respectively.

[0059] From Figure 5Several patterns can be observed. After chemotherapy, ΔD in 13 patients was consistent with the response measured by the percentage of residual surviving tumor on pathological images. However, ΔV in 8 patients was consistent with the response assessed by pathology. Conversely, with immunotherapy combined with chemotherapy, only 10 patients had ΔD consistent with the response measured by the percentage of residual surviving tumor on pathological images, while 15 patients had ΔV consistent with the response assessed by pathology. This suggests that after chemotherapy, ΔD is a better predictor of CT RECIST response than ΔV in the early stages of treatment. However, after immunotherapy combined with chemotherapy, ΔV is a better predictor of CT RECIST response than ΔD in the early stages of treatment.

[0060] It is worth mentioning that we started from Figure 5 The third line of the study observes that when using ΔV to predict CT RECIST responses to two types of treatment, some patients lacking treatment response were misclassified into the category of those with good treatment response. This introduces the problem of over-evaluation and could lead to serious clinical misdiagnosis. We hypothesize that the problem of over-evaluation might stem from defining CT RECIST response using a -30% threshold based on RECIST criteria. Therefore, to avoid over-evaluation, we seek a more appropriate threshold for ΔV. Specifically, by iterating through all ΔV values, we select the value that results in the highest number of patients with CT RECIST responses consistent with pathological responses as the threshold. The threshold selection criteria are as follows: Figure 5 As shown in the last line, we tentatively set the threshold for ΔV to -0.6868.

[0061] Setting a threshold for ΔV can avoid over-evaluation, such as... Figure 5 As shown in the last line, only 2 patients exhibited overassessment issues after chemotherapy, and 1 patient after immunotherapy combined with chemotherapy. Furthermore, we observed that for neoadjuvant immunotherapy combined with chemotherapy, ΔV remained more precise than ΔD, with ΔV consistent with pathological assessment in 16 out of 20 patients. Additionally, for chemotherapy, ΔV is now closer to pathological assessment response than before, with ΔV consistent with pathological assessment in 12 out of 15 patients.

[0062] We then modeled the relationship between the three continuous variables using linear regression, as shown in Table 4. Specifically, we calculated the Pearson correlation coefficient (PCC), p-value, and R-squared between ΔD and ΔV (percentage of residual surviving tumor). Linear regression analysis was performed on all 35 patients, with separate analyses based on treatment type. For all patients, we observed that ΔV was superior to ΔD in both PCC, p-value, and R-squared for the percentage of residual surviving tumor. Similar patterns were observed for different treatment types. This suggests that ΔV can be more accurate than ΔD in predicting prognosis at an early stage of treatment.

[0063] Table 4: Linear regression results for 35 patients

[0064]

[0065]

[0066] 8 Survival analysis.

[0067] To further validate that AV is a better CT RECIST marker than AD in the early treatment phase, we selected 47 patients in the late phase and recorded their overall survival. The characteristics of the 47 patients in the late phase are shown in Table 2. We performed survival analysis and showed the Kaplan-Meier survival curves to divide the patients into two groups with high or low survival rate according to AV and AD respectively. Then, we compared the two groups for the difference in survival rate with AD and AV as the criteria. Figure 6 The results of survival analysis for 47 patients in the late phase are shown. The left panel shows the survival analysis using variable AD, and the other panel uses AV. We used the threshold suggested in RECIST -30% to analyze AV and AD respectively. The blue line here represents AV and AD greater than -30%, and the red line represents AV and AD less than -30%. The y-axis represents the proportion of survival, and the x-axis represents OS. We used the log-rank test to compare the two groups in each panel.

[0068] As shown in Figure 6 the survival curves in the left panel cannot be distinguished, and the p-value of 0.68 indicates that there is no significant difference in OS for patients grouped by AD. In contrast, in the right panel, the survival curves can be well distinguished, and the p-value of 0.0093 indicates that there is a significant difference in OS between patients grouped by AV. These findings indicate that, compared to AD, AV can more accurately indicate those patients with high long-term OS.

[0069] Finally, our study relies on AI-based methods to segment lung CT images and compute tumor volume. Previously, CT RECIST response relied on manually computed diameter changes. In fact, it is very difficult for clinicians to obtain tumor volume by hand. With the development of artificial intelligence, tumor lesion segmentation technology enables clinicians to automatically estimate tumor volume. In addition, it also facilitates the application of CT RECIST response measured by tumor volume change. In addition, in our work, clinicians need to check the accuracy of segmented lesions. In order to improve their efficiency, we plan to improve the accuracy of lesion segmentation results, especially for those challenging lesions. For example, based on existing image segmentation methods, lesions that closely adhere to the chest wall and lesions located near the bronchus are usually difficult to separate from normal organs. In addition, the interaction between the segmentation model and the clinician should be strengthened. For example, a platform can be provided for clinicians to easily modify the segmentation results generated by artificial intelligence. Then, the AI-based segmentation model should try to learn from the clinicians' modifications to improve its segmentation ability.

[0070] In addition, it should be understood that although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that those skilled in the art can understand. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

Claims

1. A diagnostic device, characterized in that: The diagnostic device includes a CT scanner, and has a module that calculates the rate of change in tumor volume before and after early treatment of lung cancer based on CT diagnostic results. This rate of change in tumor volume before and after early treatment of lung cancer is a prognostic marker used for neoadjuvant immunotherapy and chemotherapy for lung cancer. The diagnostic device includes the following modules: (1) Lesion segmentation module, wherein the lesion segmentation module is based on the lung cancer CT image data of the Medical Segmentation Decathlon dataset to pre-train the NNU-net deep model framework, and the pre-trained model is used for CT image data segmentation; (2) Lesion optimization module: The lesion optimization module uses the Hounsfield Units thresholding method to remove pixels with Hounsfield Units values ​​less than -100 from the lesions obtained based on NNU-net segmentation, and obtains the main lesion with the largest connected component. (3) Lesion volume calculation module, wherein the lesion volume calculation module uses open-source pyradiomics to obtain the diameter and volume of the tumor, wherein the tumor diameter is defined as the maximum pairwise Euclidean distance between the vertices of the tumor surface grid, and the tumor volume is calculated by the Marching Cubes algorithm, wherein the calculation of the tumor volume includes the following steps: based on a Cube Search A triangular mesh, and for each triangular mesh, a tetrahedron is constructed. Its volume The following calculation formula is used to obtain... ， in , as well as This refers to the length of the three sides that make up the triangular mesh; the final tumor volume value is obtained by summing the volumes of all tetrahedrons using the summation formula.

2. The diagnostic device according to claim 1, characterized in that: The lesion segmentation module performs min-max normalization on the training data, normalizing its data range to 0-255; then, it uses the normalized dataset to fine-tune the parameters of the pre-trained NNU-net model; finally, it applies the fine-tuned model to a JPG format lung cancer dataset to obtain the corresponding lesion segmentation results.

Images