The research process of this study is shown in Figures 1 , 2 . The research process comprised three steps: (1) image segmentation; (2) image feature extraction and model construction; and (3) nomogram model construction using clinical features and imaging omics.

The study included patients from Qilu Hospital of Shandong University and Shandong Provincial Hospital who were hospitalized between January 1, 2017, and December 30, 2020. All patients were diagnosed with HBV-HCC and underwent percutaneous microwave ablation. The inclusion criteria are as follows: (1) patients in whom HBV-HCC was clearly diagnosed; (2) without previous treatment related to HC; (3) MRI was performed before treatment; and (4) absence of other tumors. HBV-HCC was diagnosed according to the 2018 Practice Guidance by the American Association for the Study of Liver Diseases (AASLD) (Marrero et al., 2018). The exclusion criteria are as follows: (1) patients who had received other treatments, such as transcatheter arterial chemoembolization (TACE); (2) patients with incomplete information; (3) those that were lost to follow-up; and (4) cases with poor MRI images. The specific process is shown in Figure 3 .

The primary endpoint was early recurrence (ER), defined as the time from the start of treatment to recurrence within 1 year. This retrospective study was approved by the Ethics Committee of Shandong University Qilu Hospital (No. KYLL-2022(ZM)-961) and Shandong Provincial Hospital (SWYX: No. 2022-079), and the requirement for written informed consent was waived.

All patients were scanned using MRI (1.5T, MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany) before surgery using a 16-channel abdominal coil to receive signals. The patient was placed in the supine position. The head was inserted into the machine first, and the examination scope covered the entire liver. A T2WI coronal scan was performed, followed by an axial breath-hold fat suppression sequence T2WI (scan parameters: repetition time (TR), 3000–4000 ms; echo time (TE), 90–104 ms; slice thickness, 5 mm; slice interval, 0.5 mm). Dynamic contrast-enhanced scanning used axial breath-hold volume interpolated breath-hold examination (VIBE) (TR: 3.92 ms, TE: 1.39 ms, flip angle: 9°, slice thickness: 5 mm). The acquisition included T2-weighted imaging (T2WI), T1-weighted imaging (T1WI), and diffusion-weighted imaging (DWI). Images of the arterial phase (AP), portal venous phase (PVP), and delayed phase (DP) were obtained at 35 s, 50–65 s, and 5–8 min after the injection of the contrast agent. We collected the required scanning sequence AP, DP, T2WI, and DWI. Furthermore, AP, DP, T2WI, and DWI scan images of the patient’s MRI were obtained from the PACS system (DICOM format) and uploaded to ITK-SNAP software (Version 4.0.0-alpha-3). The radiologist delineated the region of interest (ROI) on the four-phase images along the lesion edges. The ROI mapping was performed independently by two radiologists with 6 years (physician 1) and 8 years (physician 2) of experience in abdominal imaging diagnosis, respectively. The radiologists were blinded to the purpose and content of the study. If the two physicians disagreed, they were asked to discuss the areas of disagreement and arrive at a consensus.

Before extracting the radiomics features, the AP, PVP, and DP images were preprocessed. Linear interpolation was used to resample the images to 1 mm × 1 mm × 1 mm to attempt to reduce the influence of different layer thicknesses. Afterwards, gray-level discretization processing was applied to convert the continuous images into discrete integer values to enhance the robustness of the imaging features (Wu C. et al., 2022). The inputs of the deep learning model had to be fixed-size images containing the corresponding tumor area; therefore, the MRI images were cropped using a 3D bounding box covering the tumor region. The tumor image was then scaled to a voxel size of 64 × 64 × 32 as the input of the model. For each MRI sequence, we built a 3D ResNet50 model to extract the deep learning-based radiomics (DLR) features. Afterwards, the DLR features were passed into a transformer network using multi-headed self-attention that considered each DLR feature as a token and related each element to every other element. Specifically, the image features (fDP, fAP, fDWI, and fT2WI) of the four sequences were concatenated to obtain Ftrans. After linear projections, the concatenated features are mapped to the key (K) vector, query (Q) vector, and value (V) vector, and a self-attention layer computed a query-key product as described below:

The correlation was calculated by the dot product between the Q vector and the K vector, and the attention score value was scaled to 0–1 using the Softmax operation and multiplied with V to obtain the optimized feature vector. A multi-head self-attention mechanism was then applied, and the outputs were fused in a weighted manner. After processing through the Transformer module, we obtain the image feature representation Ftrans, which contained fDP, fAP, fDWI, and fT2WI, and acted as the final feature of the image. Finally, we combined the feature connections and used a fully connected layer to compute the DLR score.

Logistic regression analysis was used to screen for potential clinical factors that may influence early recurrence after microwave ablation, such as sex, age, hepatitis B surface antigen (HBsAg), alpha-fetoprotein (AFP), alanine aminotransferase (ALT), aspartate aminotransferase (AST), gamma-glutamyl transferase (GGT), alkaline phosphatase (ALP), total bilirubin (TBIL), albumin (ALB), hypersplenotrophy, hypertension, tumor number, and tumor size. A combined model integrating the DLR score with independent clinical factors was subsequently established via multivariate logistic regression to enhance predictive accuracy. To facilitate clinical application, a nomogram was constructed based on the combined model, visualizing the contributions of each predictor through a graphical scoring system that translates individual patient data into a personalized probability of early recurrence within 1 year.

Clinical features were analyzed using the chi-square test or Fisher’s exact test, as appropriate. Univariate logistic regression analysis was performed to identify clinical factors associated with early recurrence after microwave ablation, with variables having p-values < 0.05 considered for inclusion in multivariate logistic regression to construct the clinical model. The deep learning-based radiomics features were extracted and fused as detailed in the methods section, culminating in the computation of the DLR score. Binary logistic regression was employed to develop the radiomics model, clinical model, and combined nomogram integrating the DLR score with selected clinical predictors. Model discriminatory performance was evaluated using the operating characteristic curve (ROC), calculating the area under the curve (AUC) with 95% confidence intervals (CI), along with sensitivity, specificity, accuracy, F1-score, positive predictive value (PPV), and negative predictive value (NPV). Model calibration was assessed using calibration curves. Clinical utility was determined using DCA to quantify net benefits across various threshold probabilities. All statistical analyses were performed using R software version 4.1.0 (R Foundation for Statistical Computing, Vienna, Austria), SPSS version 26.0 (IBM Corp., Armonk, NY, USA), and MedCalc version 20.010 (MedCalc Software Ltd., Ostend, Belgium).

A total of 166 individuals were included in this study, including 116 in the training group and 50 in the validation group. There were 70 patients (42.2%) with HBV-HCC recurrence within 1 year after microwave ablation. A total of 116 patients were included in the training group, 55 (47.4%) of whom had recurrence within 1 year, while 15 (30.0%) of the 50 patients included in the validation group also had recurrence within 1 year. The clinicopathological features of the training group and the validation group are shown in Table 1 ; the differences between the two groups were not statistically significant (P < 0.05).

The ROC curve was used to evaluate the efficacy of the imaging model in predicting HBV-HCC recurrence within 1 year after microwave ablation. The DLR model we established in the training group had an AUC of 0.847 (95% CI: 0.768–0.907), a sensitivity of 94.5% (95% CI: 0.851–0.981), a specificity of 68.8% (95% CI: 0.564–0.791), an accuracy of 81.0%, an F1-score of 82.5%, a positive predictive value (PPV) of 73.3%, and a negative predictive value (NPV) of 93.3% in the training group ( Figure 4A , Table 2 ). In the validation group, the AUC was 0.779 (95% CI: 0.639–0.884), sensitivity was 73.3% (95% CI: 0.480–0.891), specificity was 77.1% (95% CI: 0.610–0.879), accuracy was 76.0%, F1-Score was 64.7%, PPV was 57.9%, and NPV was 87.1% ( Figure 4B , Table 3 ).

Univariate logistic regression analysis was performed to identify clinical factors associated with early recurrence after microwave ablation; variables with p-values < 0.05, such as AST levels and portal hypertension, were considered for inclusion in the multivariate logistic regression to construct the clinical model ( Table 4 ). A clinical prediction model was established with an AUC of 0.681 (95% CI: 0.588–0.765) and a sensitivity of 65.5% (95% CI: 0.523–0.766), a specificity of 67.2% (95% CI: 0.547–0.777), an accuracy of 66.3%, a F1-Score of 64.9%, a PPV of 64.3%, and an NPV of 68.3% ( Figure 5a , Table 2 ). Similarly, in the validation group, the clinical model was established with an AUC of 0.771 (95% CI: 0.631–0.878), sensitivity of 60.0% (95% CI: 0.357–0.802), specificity of 88.6% (95% CI: 0.740–0.955), accuracy of 80.0%, F1-Score of 64.2%, PPV of 69.2%, and NPV of 83.7%. ( Figure 5c , Table 3 ).

Afterwards, we used the deep learning DLR score, AST, and portal hypertension as factors for the multivariate logistic regression analysis to build personalized prediction models. According to the results of the multivariate logistic analysis, two independent clinical factors were combined with DLR scores to establish a clinical-radiological model. The combined model showed the best evaluation efficacy in both cohorts, with AUC, sensitivity, and specificity in the training group being 0.868 (95% CI: 0.793–0.924), 96.3% (95% CI: 0.877–0.990), and 67.2% (95% CI: 0.547–0.777), respectively ( Figure 5b , Table 2 ). The AUC, sensitivity, and specificity of the validation group were 0.842 (95% CI: 0.711–0.930), 80.0% (95% CI: 0.548–0.930), and 83.8% (95% CI: 0.673-0.919), respectively ( Figure 5d , Table 3 ). In terms of accuracy, the nomogram also outperformed other models (training group, 81.0% vs. DLR: 81.0%, clinical: 66.3%; validation group, 82.0% vs. DLR: 76.0%, clinical: 80.0%). In addition, the F1-score was highest for the nomogram (training group, 0.828, vs. DLR: 0.825, clinical: 0.649; validation group, 0.727 vs. DLR: 0.647, clinical: 0.642), reflecting its balanced performance in imbalanced data scenarios. Regarding negative predictions, the nomogram also demonstrated the best NPV in both the training group (95.3% vs. DLR: 93.3%, clinical: 68.3%) and the validation group (90.1% vs. DLR: 87.1%, clinical: 83.7%), indicating high reliability in ruling out the target condition. The nomogram emerged as the most robust tool, combining high discriminative power (AUC) with balanced sensitivity-specificity profiles. Finally, we visualized the model as a nomogram for clinicians ( Figure 5e ).

A calibration curve and the Hosmer-Lemeshow goodness-of-fit test showed that the predicted value and the actual result had a good fit in the training and validation groups ( Figures 6a, b ). We used decision curve analysis (DCA) to evaluate whether this nomogram could contribute to clinical treatment strategies. The DCA curves ( Figures 6c, d ) showed that the nomogram had a good net benefit in predicting HBV-HCC recurrence within 1 year after microwave ablation at most threshold ranges of 0.1–1.0. This indicates that patients with HBV-HCC can benefit immensely from the use of this nomogram.

Our study has some limitations. First, this was a retrospective study. Prospective studies have better control of confounding factors and bias; therefore, we aim to design prospective studies in the future to further verify our conclusions. Second, while our external validation cohort (n=50) was well-characterized, its modest size and single-institution origin may limit its generalizability across diverse populations or imaging protocols. Multicenter validation is essential to confirm robustness, particularly for rare subtypes or borderline cases. Preliminary cross-institutional data (not included here) also suggest reproducible performance, supporting the scalability of our approach. To address this, we have initiated a prospective multicenter trial to pool data from heterogeneous sources, ensuring broader applicability. Beyond validation, seamless clinical integration is critical. We aim to develop an implementation framework to embed the nomogram as a DICOM-compatible tool within PACS, enabling automated risk score generation alongside radiology reports. Feedback from clinicians (e.g., via Likert-scale surveys) will guide usability refinements, ensuring practical utility. Regulatory considerations will also be addressed for clinical deployment. Third, in our study, the deep learning model was designed to process the ROI that encompasses the tumor, as annotated by the radiologists, allowing for image features to be learned effectively from the ROI. A key advantage of employing deep learning in radiomics analysis is its robustness to variations in boundary delineation. While inter-observer variability was not quantitatively assessed (e.g., via Dice similarity coefficient or Kappa statistics) due to the model’s resilience to such variations, future studies could incorporate such evaluations for enhanced reproducibility. Additionally, we deliberately excluded the portal venous phase (PVP) due to inconsistent image quality and a lack of standardization across retrospective datasets, thereby ensuring model reliability and generalizability. Despite this omission, our model achieved strong performance (AUCs 0.868–0.842) by effectively leveraging AP, DP, T2WI, and DWI sequences, demonstrating their sufficiency for recurrence prediction. While we acknowledge the clinical value of PVP, its exclusion was justified in this real-world study; however, in future prospective studies, we will assess its incremental benefit under standardized protocols. Finally, in China, most HCC cases were caused by HBV infection; therefore, the participants included in this study were patients with HBV-HCC. Since the prognosis of HCC is different due to different etiologies, the model needs to be further validated in patients with HCC due to other etiologies prior to its large-scale application.