An endorectal ultrasound-based radiomics signature for preoperative prediction of lymphovascular invasion of rectal cancer

According to global cancer statistics in 2020, rectal cancer (RC) is a major threat to humans. The RC has the third-highest recorded incidence of cancer, which is listed after lung cancer and breast cancer, and the cancer-related mortality rate of RC is ranked second after lung cancer [1]. Early detection and diagnosis is crucial to effectively treat RC. Unfortunately, most patients are diagnosed with locally advanced rectal cancer (LARC), which means RC with stage T3 or T4 [2]. The standard treatment option for LARC is neoadjuvant chemoradiotherapy followed by total mesorectal excision (TME) [3, 4]. However, even though patients with LARC have already received neoadjuvant chemoradiotherapy, the 5-year overall survival rate is only 66.3% [5].

The poor prognosis of RC is related to various factors, including lesion size, stage, lymphovascular invasion (LVI) status, tumor marker status, immune indicators that are positive, RC’s sensitivity to radiotherapy and chemotherapy, and so on [6].

Previous studies have shown that vascular invasion is an independent risk factor for poor prognosis of RC [7, 8]. Now, the LVI status can be obtained by only postoperative pathology, which therefore cannot guide whether the patient should be treated more aggressively preoperatively. Therefore, preoperative prediction of LVI in RC is helpful for clinical decision-making.

There are numerous means of preoperative examination for RC. Colonoscopy can observe the general surface of the mass and perform the biopsy; however, the disadvantage is that the depth of infiltration and extraintestinal invasion could not be observed. Magnetic resonance imaging (MRI), especially high-resolution MRI, is the first choice for preoperative imaging diagnosis and postoperative follow-up review, which can accurately perform tumor node matastasis (TNM) staging, but its high cost makes it difficult to repeat multiple examinations, and it is always difficult to distinguish the T2 stage from the T3 stage [9, 10]. Endoscopic ultrasound (EUC) and transrectal ultrasound (TRUS) are cheap and repeatable and can visualize structures at multiple levels of the intestinal wall, but they still do not solve the problem of the preoperative prediction of LVI [11, 12]. To sum up, to optimize individualized clinical treatment, a method that can noninvasively and preoperatively predict the presence or absence of LVI is an urgent clinical need.

In recent years, radiomics has been introduced and has the potential to provide help for the above clinical challenges. Radiomics, using computers to excavate the data of medical images (ultrasound, CT, MR, etc.) to discover the intrinsic relationship between medical images and features, such as diagnostic localization, biological behavior, and treatment outcome of diseases, could provide a reference for clinical decision-making. Several recent reviews systematically reviewed artificial intelligence (AI) models and machine learning about radiomics based on high-resolution MRI which could be used to predict the response of chemotherapy and distant metastasis of RC, and the results showed that this novel imaging technology might be helpful to clinical decision-making [13‐16]. Chinese scholars have reported that CT radiomics can predict whether lymph node metastasis exists preoperatively in RC [17], and CT radiomics and MR radiomics can predict colorectal cancer LVI [18, 19]. Also, high-resolution MR radiomics could predict peripheral nerve invasion [20] and disease-free survival [21] in RC, assessing response to treatment for RC [22, 23]. Ultrasound radiomics can predict subtypes of liver cancer [24] and determine whether lymph node metastasis exists in thyroid cancer [25]. However, to the best of our knowledge, studies of endorectal ultrasound-based radiomics in the LVI have not yet been reported. Our study aims to utilize endorectal ultrasound-based radiomics techniques to predict LVI preoperatively and noninvasively to aid in making clinical decisions.

RC has a high incidence and poor prognosis. Microvascular invasion is an important factor in the residual and recurrence of RC [26]. At this stage of clinical practice, microvascular invasion is mainly identified by postoperative pathology, with a certain lag. Our study purpose was to develop a preoperative prediction model for LVI of RC using endorectal ultrasound-based radiomics. The model showed good predictive value in both training and validation sets, with AUC of 0.849 and 0.781, respectively, suggesting that endorectal ultrasound-based radiomics may predict the vascular invasion of RC to provide a reference for clinical treatment decisions.

Radiomics is useful in the diagnosis, staging, treatment response, and prognosis of many solid tumors. For example, radiomics can be applied to the differential diagnosis of HCC [24], the assessment of the efficacy of neoadjuvant chemotherapy for breast cancer [27], and the characterization, grading, and prediction of tumor outcome in kidney cancer [28]. Radiomics also has important value in diagnostic staging [29], treatment response [23, 30], and prognosis of RC [21]. For LVI of RC, CT-based radiomics and MRI-based radiomics have good diagnostic efficacy for microvascular infiltration of RC [18, 19]. However, endorectal ultrasound-based radiomics has not been reported for vascular invasion in RC.

Ultrasound is an inexpensive, quick, and noninvasive examination, which reflects its worth in offering TNM stage information of RC. Patients may gain benefits if this examination combining radiomics is used to predict the LVI of RC noninvasively before resection.

There have been some studies of radiomics predicting vascular invasion in RC. Zhang et al. [19]. applied MR radiomics to predict LVI with an AUC of 0.884 in the training set and 0.876 in the validation set, but they included only 94 patients in this study, and this study extracted only 1188 radiomic features. A study using CT radiomics to predict RC LVI also extracted only 396 radiomic features [18]. Compared with the above two studies, our study had more patients and features, including a total of 203 patients and extracting 5936 radiomic features.

Nevertheless, radiomics has limitations, especially in its operator dependence on manual outlining of the ROIs and the generalizability of the model [31]. For example, Yao et al. [32]. predicted HCC subtypes, KI-67, and microvascular invasion (MVI) by using radiomics of multimodal ultrasound images; however, they only performed a model of the training set without the validation set and did not perform the ICC test. Li et al. [33] applied CT radiomics to develop a model for predicting distant metastasis in RC; in this study, the validation set was available, but the ICC test was not performed. The above studies did not perform ICC tests on the ROIs to assess the consistency between different operators and the reproducibility of these models. In our research, the prediction model is not only established by a training set but also verified by a validation set to ensure the universality of the model. At the same time, we performed an ICC test on the outlined portion of the ROIs to test the consistency between different operators. The results suggested that the ICC is above 0.75 in 90.1% (5350/5936), which guaranteed the reproducibility of our study, and that similar results can be obtained when the model is used by different analysts. Also, in the process of radiomics performance, we used CLAIM (Checklist for Artificial Intelligence in Medical Imaging) as the Radiomics Quality Score to make sure the methodological quality [34]. And this checklist was added in the Additional file 1.

This study has the following limitations. First, the case selection bias may exist in our retrospective study, which should be avoided by prospective studies in the future. Second, the ultrasound images used in this study were from different operators using different machines, and these confounding factors could potentially affect the predictive performance of the model, so the next step requires a prospective cohort study with strictly controlled variables (e.g., one operator, with one machine to acquire images). Third, ultrasound images only take the largest section for analysis. Although it can represent a lesion, it still cannot represent the whole lesion, and information may still be missed. Future studies need to analyze the overall lesions by forming three-dimensional images. Four, the limit number of cases is also one of our limitations. We need further research with large samples and multicenter data in the future.

In summary, we have developed a prediction model for endorectal ultrasound-based radiomics, which is expected to be used to predict the presence of RC LVI. This method can be promoted in clinical practice because it can noninvasively predict the presence or absence of LVI before surgery and thus influence clinical decision-making.