A virtual biopsy study of microsatellite instability in gastric cancer based on deep learning radiomics

Gastric cancer (GC) is a highly heterogeneous malignancy caused by multiple factors and is a global public health problem. The incidence of GC varies according to geographical location and is particularly high in Asia (age-standardized [global] incidence: 32.5 per 100,000 men; 13.2 per 100,000 women) [1, 2]. MSI-positive GC is one of the major molecular subtypes of GC as defined by the Cancer Genome Atlas Group and accounts for 10–22% of all GC patients [3]. The deletion of any of the mismatch gene repair proteins (MLH1, PMS2, MSH2, and MSH6) leads to microsatellite instability (MSI/MSI-H) [4, 5]. In recent years, immune checkpoint inhibitors have shown great potential in the treatment of progressive GC [6]. MSI is an important predictive biomarker for evaluating the effect of anti-programmed cell death-1 (PD-1) immunotherapy in GC patients. Several clinical trials have confirmed that the objective response rate and survival were significantly better in the MSI group than in the microsatellite stabilization (MSS) group in anti-PD-1 immunotherapy for advanced GC [7‐9]. However, there is a large individual variation in the efficacy of anti-PD-1 therapy [10], so it is important to select patients who are most likely to benefit. Currently, immunohistochemical staining (IHC) and PCR molecular testing are mainly used to detect MSI expression levels [5, 11, 12]. Given the spatial heterogeneity of MSI expression, a small piece of tissue obtained by invasive biopsy may not be sufficiently representative of the entire tumor region [13], thus affecting the assessment of MMR protein expression. Although universal testing for MSI has been recommended in the NCCN guidelines for patients with GC [14], many patients are not tested due to the invasive, time-consuming and expensive nature of tissue biopsy. Therefore, there is an increasing need to develop a non-invasive method for the holistic assessment of MSI expression in GC.

With the development of computer-aided medicine, accurate assessment of tumor’s pathological features has been achieved by combining radiomics and deep learning to extract and analyze quantitative radiological features, known as 'virtual biopsies,' which provide a reference standard for conventional biopsies [15‐17]. Several studies have demonstrated the potential role of deep-learning-based radiomics in predicting lymph node metastasis [18] or response to neoadjuvant chemotherapy [19, 20] in GC, so this study explored the non-invasive assessment of GC biomarkers based on such methods using preoperative computed enhanced tomography (CECT) images and clinical data. It was also visualized as a nomogram to evaluate the potential application as a virtual biopsy tool in clinical auxiliary diagnosis.

We used clinical information and pretreatment CECT images of 223 GC patients from our medical center to develop a virtual biopsy model supported by deep learning algorithms. To the best of our knowledge, we are the first to apply deep learning algorithms to the optimization of radiomics feature building in the GC MSI radiomics. Clinical imaging models supported by MLP effectively improve the ability to identify MSI/MSI-H in GC patients (AUC: training set 0.883, testing set 0.802). Accurate MSI identification is critical for individualized systemic treatment of GC patients, which can benefit patients receiving anti-PD-1 immunotherapy and improve GC patient survival [9, 24, 25].

All valid samples in a unit period were covered in this study, avoiding selection bias as much as possible. The incidence of MSI in GC patients in the study was 18.4%, which is consistent with the current Global epidemiology on GC-related MSI (10–22%) [3, 5, 26]. LASSO regression analysis of clinical baseline data found that age, sex, and clinical T stage were independent clinical risk factors closely related to MSI expression in GC patients. Considering the relationship between advanced age [27‐30] and female sex [28, 30], GC patients and MSI are consistent with previous related studies. Current evidence suggests that the MSI-H phenotype in patients with sporadic GC is closely associated with hypermethylation of the promoter CpG island causing silencing of the hMLH1 gene, which leads to progressive loss of MLH1 protein expression [31‐33]. Similar findings were found in our recruited patients, with more than 83% (n = 35) of MSI-expressing patients having a deletion in the expression of the mismatch repair protein MLH1. Interestingly, researchers such as Nakajima et al. [34] and Kim et al. [35] found that the methylation of the mismatch repair gene hMLH1 was age dependent, and its incidence was positively correlated with age. hMLH1 methylation is more common in elderly gastric cancer patients, which seems to explain why MSI/MSI-H was more common in older GC patients in our study. Compared with MSS/MSI-L, GC with the MSI/MSI-H phenotype is less aggressive, representing a better prognosis in the early stage of GC [27, 36]. At the molecular level, recent studies have found that changes in the genetic and epigenetic characteristics of the GC genome often occur in the early stages of the tumor [37, 38], which supports that hMLH1 promoter methylation-induced MSI-H is more likely to be seen in the early stage (TNM stages I–II) presumed in GC patients, thus defining MSI as an early molecular event of GC. This is also confirmed in the reports of Polom et al. [30] and Jahng et al. [39]. However, we did not observe the previously demonstrated significant relationship between tumor location [29, 30, 36] and MSI, which may be related to differences in the patient population or regional prevalence of our recruitment. Taken together, these potentially characteristic clinical factors reflect higher histological heterogeneity in MSI tumors.

The development of radiomics technology makes it possible to capture the heterogeneity of tumor molecules in clinical images, which can provide objective and quantitative support for cancer molecular biological detection and personalized treatment [40]. We finally selected 15 radiomics features (including 1 first-order feature, 1 shape-based feature, and 13 filtered features) from the filtered and unfiltered images to build the model. The first-order feature describes the distribution of voxel intensities in CT images, and the shape-based feature is a description of the two-dimensional tumor regions’ size and shape. The filtered features are first-order statistical features and texture features extracted from the filtered images and reconstructed by transforming the Laplacian of the Gaussian spatial bandpass filter or wavelet filter, where the texture features include the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix (GLRLM), gray-level size zone matrix (GLSZM), and gray-level dependence matrix (GLDM). With these two different filtering strategies, the specific structure of the original image is enhanced. These features reflect the differences in spatial morphology, pixel intensity, and texture of tumor regions and may represent spatial and temporal heterogeneity in tumor tissue and characteristic biological phenotypes [23, 41]. This may explain the ability of the Rad-score established in this study to exhibit differentiated GC MSI in both cohorts. Additionally, accurate and efficient tumor region segmentation methods are important to ensure the quality of quantitative image features. It has been confirmed that 3D Slicer has better segmentation algorithms and higher segmentation accuracy than other image segmentation software while being more accessible as free open-source software [42]. Moreover, the radiomics features extracted after image segmentation using 3D Slicer have higher robustness and are, therefore, recommended for use in high-throughput data mining efforts for medical oncology imaging [43].

The exploration of radiomics in the field of GC biomarkers has started. Li et al. [44] used radiomics to construct a predictive model for detecting GC HER-2 expression, with an AUC of 0.799 [95% CI: 0.704 − 0.894]. Based on the information of 189 patients, Liang et al. [45] first constructed a predictive model based on logistic regression analysis to explore the feasibility of predicting GC-related MSI status, and the AUC was 0.8228 [95% CI: 0.7355–0.9101]. However, traditional radiomic method brings challenges in image segmentation, standardization, acquisition, and reconstruction. As an emerging means of quantitative image analysis, deep learning can optimize such limitations and improve the accuracy and reliability of prediction models [16, 46, 47]. The combination of deep learning and radiomics has shown promising results [18‐20, 48]. In this study, we first attempted to apply deep learning algorithms to the calculation and reconstruction of GC-related MSI radiomic features. Compared with previous [45] research, our results were encouraging and obtained a higher AUC (0.883 and 0.802). This may be attributed in the optimization of radiomic features by deep learning algorithm, the increase in sample size, and the level of image segmentation in our study. Additionally, we tried to add pathological features to the combined model, but the final result did not significantly improve the model's predictive performance, reflecting the independent value of pretreatment radiomics in predicting MSI status in GC patients. To increase clinical practicability, we visualized the clinical image model into a nomogram based on logistic regression to generate the prediction probability of MSI so that it could be used as a virtual biopsy tool to support clinical medical decision making.

It should be noted that our study also has limitations. First, this is a single-center retrospective study, and inevitably, there is a patient selection bias. Although we tried to include a relatively more number of cases (n = 223), the study sample is still small, considering the high incidence of gastric cancer in Asia, which may affect the generalizability of the model. Second, due to the inherent black box property of machine learning, the process between model data input and output is difficult to interpret, and this lack of transparency has implications for clinical practice, so we plan to apply more transparent and interpretable medical algorithms in future [49, 50]. Finally, we also noted the place of dual-energy CT (DECT) in radiomics, and in future, we plan to collect cases in DECT centers to study the benefit of DECT in GC MSI virtual biopsy studies.