In our study we developed different models to predict response of target lesions in patients with HCC who underwent radiation segmentectomy using baseline demographic and clinical data as well as radiomics from baseline MR images. The model using a combination of demographic/clinical data and radiomics from the PVP (Model D) showed the best performance with fair prediction power for CR at 6 months which was similar to the performance of using radiomics alone (Model C). On the other hand, baseline demographics/clinical parameters alone (Model A) and AFP alone (Model B) showed significantly worse performance compared to the models using radiomics (Models C, D, and E).

Quantitative studies evaluating HCC response prediction to radiation segmentectomy from MRI baseline images are sparse. To the best of our knowledge there are, to date, only a few studies that evaluated pre-radiation segmentectomy imaging findings in HCC patients using CT [22], FDG-PET/CT [37], SPECT CT [38] or MRI [21] to predict response and outcome, respectively. The study by Reiner et al. [22] found a sensitivity of 88% and a specificity of 75% for distinguishing responders from non-responders after radiation segmentectomy using histogram features quantification from baseline perfusion CT images [22] which was higher compared to our results (Model D, 70.6% and 66.2% respectively). In contrast to our study however, their study defined CR or PR as responders, whereas in our study only CR was considered as response. Notably, perfusion CT is not routinely used for radiation segmentectomy treatment planning in clinical practice. Jreige et al. [37] suggested that quantitative functional parameters from pre-radiation segmentectomy FDG-PET/CT studies such as SUVmax and tumor-to-liver uptake ratio are able to predict overall survival and progression-free survival. However, currently FDG-PET/CT plays a minor role in HCC evaluation and staging which might limit the role of functional parameters in clinical practice.

The investigation conducted by Marinelli et al. [21] yielded promising outcomes through the utilization of MRI radiomics features for prognosticating CR subsequent to radiation segmentectomy among patients with HCC, showcasing an AUC of 0.89, with a sensitivity of 80% and specificity of 88%. Notably, this performance surpassed that of baseline clinical characteristics (AUC 0.59, sensitivity 100%, specificity 28%). Nevertheless, in contrast to the methodology employed in our investigation, Marinelli et al. incorporated radiomics features extracted from both pre- and post-treatment MRI scans, a distinction that may account for the superior diagnostic efficacy observed in their study compared to ours. For our study, we opted to include only pre-TARE MRI data because we believe this approach is more practical in clinical practice compared to incorporating post-TARE images, as it would avoid subjecting the patient to an ineffective treatment and its associated potential side effects, and it may orient them towards other treatment modalities.

Another study by Chapiro et al. [39] found that in patients with HCC, the total and enhancing tumor volume from baseline MRI is associated with overall survival. Similarly, in our study a high tumor volume (represented by the feature “SHAPE_Volume”) was associated with absence of CR at 6 months after radiation segmentectomy with the highest single AUC. Our results reinforce previous results that tumor size or volume is one the most important predictors for treatment success when using intra-arterial therapies [39, 40].

A study by Hu et al. [19] found that, besides the factors such as tumor size, tumor vascularity and portal vein invasion, the number of tumors is a significant independent predictor of tumor response after chemoembolization in patients with unresectable HCC. In addition, Child–Pugh class B and C, and a high AFP value indicated poor prognosis for overall patient survival in that study. Similarly, in our study a high Child–Pugh score was associated with absence of CR and frequently selected by the prediction model with the best performance (Model D, clinical/demographics features and radiomics). Furthermore, in our study, lesions in patients with BCLC stage A had more frequently CR while, in contrast, lesions in patients with BCLC stage B, C and D showed more frequently non-CR. This may stem from the fact that the BCLC stage incorporates the number of lesions, lesion size and Child–Pugh class, which are all risk factors for treatment failure. Also, similar to the study by Hu et al. [19] we found that lower AFP at baseline was associated with CR at 6 months.

In our study, models using radiomics features performed better compared to models using demographic and clinical data alone. One explanation might be that radiomics features were extracted directly from the tumor, providing biological data specific to the target lesion. In contrast, many demographic and clinical features (e.g. age, gender, underlying liver disease, etc.) reflect broader patient-level biological data but may not directly capture the biology of the lesion itself. HCC is a complex disease with known heterogeneity in histological and molecular features, genetic alterations and oncogenic pathways influencing the disease prognosis [41, 42]. Radiomics technique can quantify tumor phenotypic heterogeneity and other imaging features that are not detectable by the human eye such as microvascular invasion [43, 44].

Current guidelines recommend that patients with early HCC should be evaluated for curative treatment options such as resection, transplantation or ablation, while arterially directed therapies (e.g. TACE or TARE/radiation segmentectomy) are indicated in patients with intermediate stage HCC who are not eligible for curative therapy or as a bridge to other curative therapies [2,3,4]. However, initial attempts to use segmental radiation segmentectomy as curative treatment for early HCC have been made [45]. Regardless of the intended treatment approach (curative or not), no prognostic score or evaluation method for response prediction after radiation segmentectomy are currently available. Therefore, our proposed model, using a combination of clinical/demographic features and radiomics may help in identifying those patients who might profit from additional or adjuvant therapy, such as immunotherapy [46].

As of now, the evidence for combining radiation segmentectomy and sorafenib or immunotherapy in patients with HCC is low [47,48,49]. However, initial results in patients with HCC showed immune activation in the local tumor microenvironment after radiation segmentectomy, suggesting that a combination of radiation segmentectomy and immunotherapy could improve patient outcomes [50]. Clinical trials have recently evaluated the combination of radiation segmentectomy and immunotherapy in patients with advanced HCC. Initial data show promising results with disease control rate of 82% and no increase of the adverse event rate [51]. However, final results of this clinical trial are not published yet.

Our study has several limitations. First, the retrospective design entails a potential selection bias. However, we made every effort in selecting consecutive patients treated with a similar technique (radiation segementectomy). Second, we did not validate our model in a separate validation set. This was mainly because the sample size of our cohort was not large enough to create an adequate training and validation sets. Therefore, we focused on constructing a robust model, using an elastic net with 100-fold cross validation for feature selection and model building. This could be peformed in a separate study. Third, in our study, radiation segmentectomy was very efficient with only a small number of lesions showing SD or even PD and these categories may therefore have been underrepresented. Fourth, we only assessed target lesion response at 6 months but did not assess long term response or survival. Fifth, we did not evaluate semantic features, as we wanted to develop methods that minimize reliance on subjective features.

In conclusion, a model consisting of radiomics features extracted from pre-radiation segmentectomy MRI and clinical parameters shows fair perforamance for predicting HCC response to radiation segmentectomy at 6 months. The performance of this model was significantly better compared to clinical parameters and serum AFP. These results need further independent validation.