Patience
This retrospective study was conducted in accordance with the ethical guidelines for clinical research of the Ministry of Health, Labor and Welfare and was approved by the Ethics Committee of Tokyo Medical University (approval number: T2019-0146). Opportunities for consent to explanation and refusal were announced on the website. Therefore, the Ethics Committee of Tokyo Medical University waived the need for informed consent.
We retrospectively reviewed the medical records of 349 patients with nonmetastatic ccRCC (T1-3N0M0) who underwent radical or partial nephrectomy at our institution between 1990 and 2008. . To develop 5- and 10-year recurrence prediction models, the investigators (AS) of this study selected a total of 131 patients based on their recurrence status and follow-up as of December 2013. Patients not followed up for 5 years were excluded. Tumors were staged according to the 2002 Union Internationale Contre le Cancer TNM classification and graded according to the Fuhrman grading system.16, 17. Pathological evaluation was performed by his two senior pathologists (MK and TN). In principle, all patients were followed up by physical examination, blood evaluation and chest radiography after 3 months and by computed tomography after 6 months. Other radiological examinations were also performed as needed. Our department recommends long-term follow-up studies as much as possible, but we do not require follow-up studies beyond 10 years.
Digital image processing for nuclear evaluation
All hematoxylin and eosin (HE)-stained slides of ccRCC tissues were digitally recorded at ×20 image magnification using a whole-slide image scanner (Nano Zoomer-RS: Hamamatsu Photonics, Hamamatsu, Japan). The pathologist selected an average of 32 her ROIs per case, excluding areas that were crushed, blurred, or heavily infiltrated with lymphocytes (Fig. 3a). A representative enlarged image of the ROI is shown in Fig. 3b. Each ROI contained fibroblasts and lymphocytes, and non-cancerous regions were manually masked (Figure 3c). Using the Ilastik software (https://www.ilastik.org), nuclear extraction was performed on the RCC image only (Fig. 3d) to create a nuclear mask image (Fig. 3e). An image of the RCC nucleus was obtained by superimposing the image in Fig. 3e on the image in Fig. 3c (Fig. 3f). At this stage, there were many nuclei still polymerizing. The final image of nuclear measurements (Fig. 3g) was obtained by overlaying an additional nuclear segmentation mask created using pix2pix (https://phillipi.github.io/pix2pix/).
Image processing to extract nuclei. (be) 9-10 regions of interest (ROI) are selected from one slide. (b) each ROI is magnified. (c) renal cell carcinoma (RCC) other than clear cell RCC is masked. (d) to extract the nucleus. (e) a mask image containing only the left nucleus is created. (f) The 3e image is superimposed on the 3c image. (g) kernel segmentation is performed using deep learning.
Extraction of quantitative nuclear morphology information
Using the CellProfiler software (https://cellprofiler.org), each nucleus was analyzed for nuclear shape-related features (such as size, contour length, circularity, maximum and minimum axis length) and chromatin texture. were evaluated with respect to the features (entropy, second angular moment) of , variance, differential moments, etc.). The following CellProfiler modules were used: object size and shape measurements, texture measurements, and object radial distribution measurements. For more information on CellProfiler morphological features, see http://cellprofiler-manual.s3.amazonaws.com/CellProfiler-3.0.0/index.html. A graphical representation of the lateralized quantitative nuclear features is shown in Supplementary Fig. S1. CellProfiler outputs 80 features per nucleus, for a total of 2,512,771 nuclei measured. Finally, we adopted the CFLCM method.18It shows nuclear heterogeneity and polymorphism across the ROI image based on the features of each nucleus. This method treats each nuclear feature as a single pixel on the image and computes the heterogeneity. Using the data output by CellProfiler, CFCLM output 960 features for each ROI and a total of 4312 her ROIs were measured.
Development and validation of recurrence prediction model using machine learning algorithm
We created two predictive models for recurrence within 5 years and within 10 years. SVM was adopted as the machine learning method. Data were analyzed using the statistical software package R version 3.6.1. I also used the package “e1071: SVM Linear Kernel”.19. First, he divided the data of 4,312 ROIs (131 cases) into four groups according to recurrence and follow-up. Group A, within 5 years of recurrence. Group B, recurrence 5-10 years. Group C, no recurrence at 5-10 years follow-up. Group D has no recurrence in his 10+ years of follow-up. The number of cases in each group was 40, 22, 37, and 32, respectively. Test data were randomly selected from each group. A total of 31 data, 10, 4, 9, and 8 in groups A, B, C, and D, respectively, were separated as test data, and the rest were used as training data for his SVM model. In the 5-year recurrence model, group A data were recurrence data and groups B, C, and D data were recurrence-free data. The number of cases was as follows: 40 recurrences (30 training, 10 testing). Recurrence-free, 91 (70 training, 21 testing; Figure 4).
Include cases in the 5-year recurrence model.
In the 10-year model, Groups A and B were the recurrence groups, and Group D was the non-recurrence group. Although we could not use the data for all cases in Group C as training data, the four cases in Group C that were included as test cases in the 5-year model were included as test cases in the 10-year model, resulting in 14 included the test of Case (Fig. 5). For both models, the total cases were randomly split into training and test sets (3:1). We used the average recurrence probability for each ROI output by the SVM as the result of the predictive model. Model accuracy was confirmed by validation of test cases for each model. Finally, in order to evaluate the timing of recurrence in the postoperative course (recurrence within 5 years after surgery, recurrence within 5-10 years, no recurrence within 10 years), test cases according to the recurrence probability calculated as follows: I made a plot of two models. For plot follow-up, we also checked the distribution of T stage, nuclear grade, and AUA risk group. Additionally, we used tracking data from December 2021 to validate the accuracy of our predictions on our test cases.
Include cases in the 10-year recurrence model.
