Radiomics in evaluation of small-hepatocellular carcinoma and benign cirrhotic nodule based on magnetic resonance plain images

Background There are adequate evidences showing that early diagnosis signicantly improves the prognosis of patients. Currently, radiology diagnosis to s-HCC is still dicult. Radiomics, a new type of quantitative imaging diagnosis method, has been widely used in the study of multi-system diseases. The purpose of this study is to evaluate the value of radiomics based on MRI plain images in differentiating s-HCC and benign cirrhotic nodule. Methods A total of 78 patients with 114 lesions (89 s-HCCs and 25 benign cirrhotic nodules) were retrospectively enrolled during September 2017 to January 2020. MRI plain images (T1WI, T2WI) of each nodule was segmented to form VOI, and 1223 quantitative radiomics features were extracted from each sequence. 10 optimal features were respectively selected from the T1WI, T2WI, and T1WI+T2WI based on SelectKBest. Logistic Regression (LR) was applied in establishing 2 radiomics models based on single sequence images (T1WI and T2WI) respectively and 1 radiomics model based on T1WI+T2WI. Clinical information (including age, gender, AFP level and the longest diameter of lesion) was collected and analyzed using binary logic analysis to obtain statistically signicant (p<0.05) clinical features. Combining the optimal radiomics features based on T1WI+T2WI and signicant clinical features, the fusion model was established using LR and was expressed in nomogram. The AUC, sensitivity and accuracy of the 4 models were obtained. Delong test was used to compare the AUC of the 4 models. Results 78 patients with 114 lesions were included in this study. There were 89 s-HCCs, 11 of nodules were e-HCCs, and the remaining 25 were benign cirrhotic nodules, including 3 DNs and 22 RNs. There was no signicant difference in diagnostic performance among the radiomics model based on T2WI, T1WI+T2WI and the fusion model. And their diagnostic performance were higher than that of the radiomics model based on T1WI. cirrhotic nodule. Radiomics features obtained from T2WI played a key role in the differentiation of s-HCC and benign cirrhotic nodule.


Background
Hepatocellular carcinoma (HCC) is the second cause of cancer-related death worldwide [1] . Since most HCCs are diagnosed in the middle and advanced stages of the disease, so that effective clinical treatment is unavailable, the median survival of less than 1 year and a 5-year survival rate of less than 10% [2] . Some clinical studies have shown that surgical intervention for s-HCC can reduce the recurrence rate and improve the cure rate of HCC patients [3][4][5] . Therefore, increasing the diagnostic accuracy of s-HCC is of great signi cance for improving the prognosis of patients.
At present, imaging has become a non-invasive and accurate method for detection and diagnosis of HCC.
Enhanced MRI provide signi cantly higher detection rate and diagnosis performance of HCC, even s-HCC, when combine with typical enhancement characteristics of HCC and hepatobiliary images obtained by hepatobiliary agents [6,7] . However, due to the incomplete angiogenesis, the manifestation of the enhancement performance of s-HCC can be atypical. Takayasu et al have found that 24% of the s-HCC between 1.1cm and 2cm in diameter can be hypoenhancement [8] . And some other studies have also pointed out that 27%-34% of s-HCC presented poor blood supply [9] . The application of hepatobiliary contrast agents is more likely to produce respiratory artifacts, which affects the image quality of late arterial phase. Moreover, as the uptake of hepatobiliary contrast agent is closely related to the liver function of patients, it can have a great impact on the images [10] . So, there are certain de ciencies in differentiating s-HCC and benign cirrhotic nodule based only on qualitative imaging features. In addition, as high-risk patients of HCC may need repeated imaging examinations during follow-up, lots of enhancement examinations will increase the cost of patients. And some patients cannot undergo enhancement examinations due to contraindications. MRI plain images is more available and can be of great use.
According to the clinical diagnosis guidelines, pathological results of high-risk patients of HCC with indeterminate nodules can be obtained by puncture biopsy. Because the sample of lesion tissue obtained by puncture is insu cient, especially cirrhotic-related nodule with strong heterogeneity, diagnostic accuracy of biopsies is relatively low. Also the procedure may cause some side effects such as tumor dissemination and bleeding in puncture tract [10]. We urgently need a more e cient and accurate noninvasive method to improve the ability to distinguish between s-HCC and benign cirrhotic nodule, so as to increase the prognosis of patients with HCC [11][12][13] .
In 2012, Lambin [14] has proposed a new image quantitative analysis tool --radiomics. By extracting and selecting quantitative imaging features, robust features can be obtained to establish objective and accurate connection between imaging and biological information. Nowadays, radiomics has been widely used in the study of in ammatory or neoplastic lesions in the brain, breast, soft tissue, and etc. In terms of liver, the researches mainly focus on the identi cation of benign and malignant lesions, the prediction of HCC microvascular in ltration, and the evaluation of postoperative e cacy of HCC. So far, there are few studies involving in differential diagnosis of s-HCC and benign cirrhotic nodule by radiomics.
The purpose of this study was to evaluate the value of radiomics based on MRI plain images in identifying s-HCC and benign cirrhotic nodule.

Study population
Patients with cirrhosis who underwent liver transplantation and hepatectomy in our hospital from January 2017 to January 2020 were retrospectively included. MRI plain and enhanced images were obtained by using hepatobiliary agents or extracellular contrast agents within 4 weeks before surgery. Ensure the consistency of all machines and scanning parameters. Enrolled patients met the inclusion criteria: the longest diameter of the lesion shown by imaging was ≤2cm; pathological diagnosis of the nodules were cirrhotic -related nodules. The exclusion criteria were as follows: patients underwent any preoperative intervention treatment; the lesion was not clearly shown on MRI plain images; the gross specimen did not correspond to the location and size of the nodules shown in the imaging. The study population was randomly divided into training group and test group at a ratio of 6:4.

MRI protocol
All enrolled patients were examined with 3.0T MRI (Verio; Siemens, Erlangen, Germany), with 8-channel body phased array coil and turbo spin-echo (TSE) sequence acquisition collection under respiratory trigger. Abdominal pressure band was applied to the lower abdomen of the patient to reduce artifacts caused by abdominal respiratory movements. The scanning sequence of MRI included: axial T1WI and axial fat-suppression T2WI. The parameters of each MRI sequence are shown in Table 1.

Image-pathology comparison method
Refer to the MRI images, the liver transplantation specimens or resected specimens were cut at an interval of nearly 1cm. Pathological specimens of the abnormal signal nodules shown in the images were corresponding to the dissected liver specimen. The pathological specimens were sectioned, dehydrated and para n-embedded, stained with hematoxylin-eosin, and then subjected to histological and immunohistochemical examination. The image-pathology comparison method was illustrated in Fig. 1  Platform. A radiologist segmented the lesions on the T1WI and T2WI to form a three-dimensional VOI; and two senior radiologists with more than 10-year experience in abdominal imaging diagnosis proofread the edge of the lesions. In case of disagreement, the lesions would be segmented again after 2 weeks (Fig. 3). All three physicians were blinded to pathological results, but were aware of the purpose and method of the study.

Extraction and selection
A total of 1223 radiomics features were extracted from the segmented images at each plain sequence, which mainly include rst-order features, 2D and 3D shape features, texture features (including gray-level co-occurrence matrix (GLCM) features, gray-level run-length matrix (GLRLM) features, gray-level size zone matrix (GLSZM) features , gray-level dependence matrix (GLDM) features, and neighbourhood gray-tone difference matrix (NGTDM) features). To avoid over tting, the dimension reduction method of SelectKBest was applied to select the optimal radiomics features.

Establishment of radiomics model
LR was applied to establish three groups of radiomics models using the 10 optimal radiomics feature selected from T1WI, T2WI, and T1WI+T2WI images, respectively. Radiomics models were evaluated using the AUC.

Collection of clinical feature
A radiologist collected clinical information of each patient, including age, gender, AFP level, and the longest diameter of lesion. The radiologist was blinded to pathological results.

Establishment of fusion Model
The clinical features with statistical signi cance (p<0.05) was obtained using binary logic analysis method. LR was applied to establish a fusion model, based on optimal radiomics features obtained by T1WI+T2WI and signi cant clinical features, and expressed in nomogram. The AUC were used to evaluate the fusion model.

Comparison
Delong test was used to compare the AUCs of T1WI, T2WI, T1WI+T2WI radiomics models and fusion model.

Statistic analysis
Statistical analysis were carried out using Medcale statistical software (http://www.medcalc.org). Binary logistic regression analysis was used to select clinical features with a p value less than 0.05. Radiomics feature extraction, dimension reduction and model establishment were obtained from the Medical Standard -Darwin Intelligent Scienti c Research Platform (Beijing, Yizhun Medical AI Co., Ltd).

Characteristics of patients
Initially, a total of 82 patients were enrolled, 4 patients were excluded due to severe ascites ; nally 78 patients with 114 lesions were included in this study. 70 patients had chronic hepatitis B cirrhosis, 3 patients had chronic hepatitis C cirrhosis, 3 patients had autoimmune cirrhosis, and 2 patients had alcoholic cirrhosis. There were 89 s-HCC, 11 of nodules were e-HCCs, and the remaining 25 were benign cirrhotic nodules, including 3 DNs and 22 RNs. Patients' detail information are shown in Table 2.

Radiomics models
The ow chart for the establishment of radiomics model is shown in Fig. 3. The optimal radiomics features obtained based on T1WI images are shown in Fig. 1 in supplementary, the radiomics scores (Rad score) are shown in Formula 1. The AUC, sensitivity and speci city of the training group were 0.757 (95% CI 0.638 -0.853), 83.02% and 66.67%, respectively. The AUC, sensitivity and speci city of the test group were 0.789 (95% CI 0.643-0.895), 88.89% and 80.00%,respectively. The receiver operating characteristic curves (ROCs) of the models are shown in Fig. 4. Formula (1) The optimal radiomics features obtained based on T2WI images are shown in Fig. 2 (2) The optimal radiomics features obtained based on T1WI+T2WI are shown in Fig. 3

Fusion model and nomogram
Binary logistic regression analysis showed that the patient's gender was statistically signi cant in differentiating between s-HCC and benign cirrhotic nodule (p = 0.023, p <0.05), and the remaining clinical features were not statistically signi cant. The nomogram of the diagnostic model obtained combined with clinical characteristics and optimal radiomics features by applying logistic regression is shown in Fig. 7. The AUC, sensitivity and speci city of the training group were 0.935 (95% CI 0.847-0.980), 94.34% and 93.33%, respectively. The AUC, sensitivity and speci city of the test group were 0.772 (95% CI 0.625-0.883), 97.22% and 60.00%, respectively. The ROCs of the models are shown in Fig. 8.

Comparison of diagnostic performance
Delong test has proved that, the diagnositic performance had no difference among radiomics model based on T2WI, T1WI+T2WI and fusion model (p 0.05). The diagnositic performances of training group of these three models were higher than that of radiomics model based on T1WI (p = 0.0379, p = 0.0213, p = 0.008).

Discussion
This study veri ed that the application of radiomics based on T1WI and T2WI can be used to differentiate s-HCC and benign cirrhotic nodule. In the radiomics model based on a single sequence, we found the radiomics features obtained from T2WI play a key role in differentiating between s-HCC and benign cirrhosis nodule. We considers this has to do with the advantages of T2WI sequence. T2WI sequence can better re ect the metabolism, blood supply and interstitial structure of lesions by showing the content of 1H. First, the rapid growth and proliferation of tumor cells, causing the increase of intracellular water molecules and phospholipid of the cell membrane, results that the content of 1H in s-HCC is signi cantly higher than that of benign cirrhotic nodule [15], which can be re ected on T2WI. Second, the growth of tumor cells can induce and promote tumor angiogenesis which in turn can facilitate the proliferation of tumor cells. Pathological studies have shown that there is little unpaired arteries and sinusoidal capillarization in RN and low grade dysplastic nodule LGDN . While in high grade dysplastic nodule (HGDN), e-HCC, s-HCC, the unpaired arteries gradually increase with the growth of lesion diameter [16][17][18]. Hidenori et al indicated that the content of tumor vessels in e-HCC was signi cantly higher than that of HGDN [19]. In summary, the heterogeneity of tumor cells and interstitial tissue structure in the lesion may cause differences of T2WI signal between s-HCC and benign cirrhotic nodule. Moreover, hypoxia caused by incomplete angiogenesis resulting in the occurrence of iron deposits in the benign cirrhotic nodule may affect the T2WI signal of lesion [20][21][22]. This study indicated that there was no statistically signi cant difference in diagnostic performance between the radiomics model based on T1WI+T2WI and T2WI single-sequence (p >0.05) , but the diagnostic performance of the two models was higher than that of the radiomics model based on T1WI single-sequence T1WI. The optimal radiomics features obtained by T1WI and T2WI through SelectKBest mainly include some rst-order features (including average value, median, and minimum in VOI) and texture features (including JointEnergy in GLCM, Dependence NonUniformity Normalized in GLDM, and Contrast in NGTDM). These texture features mainly re ect the heterogeneity of lesion images [23]. According to the radiomics score obtained in this study, the conclusion is that s-HCC has more signi cant heterogeneity than benign cirrhotic nodule. We consider that this is related to the heterogeneity of tumor cells and interstitial structures in s-HCC. As mentioned above, tumor cells in s-HCC divided and proliferated vigorously. Different genes or other macromolecule changes may appear in the daughter cells generated by multiple generations of division and proliferation, which lead to greater heterogeneity between tumor cells. There are also abundant unpaired arteries and hepatic sinusoidal tumor vessels in the tumor interstitial structure. In addition, the brous septa exists between the HCC sub-nodules or between the necrotic areas and HCC tissue [24], resulting in considerable heterogeneity in the interstitial structure of the lesion. The fusion model was established based on the combination of optimal radiomics features based on T1WI+T2WI and optimal clinical features. We expressed the fusion model in nomogram intuitively , which is more convenient for clinicians and patients to understand. The results showed that there was no statistically difference between diagnostic performance of the fusion model, radiomics model based on T1WI+T2WI and T2WI single-sequence. In the collected clinical features, gender had certain signi cance for diagnosis of HCC. This reminds male patients to be more alert to the occurrence of s-HCC. AFP is the most common and readily available  Figure 1 Images from a middle-aged male patient underwent liver transplantation, with a high-grade dysplastic nodule (HGDN) con rmed by pathology.  Images from the same patient, with moderately differentiated s-HCC con rmed by pathology. Fig. A: lesion located in segment a (indicated with black square).  Nomogram obtained by the fusion model (gender 0 denotes male)