With the development of machine learning (ML) algorithms, a growing number of predictive models have been established for predicting the therapeutic outcome of patients with hepatocellular carcinoma (HCC) after various treatment modalities. By using the different combinations of clinical and radiological variables, ML algorithms can simulate human learning to detect hidden patterns within the data and play a critical role in artificial intelligence techniques. Compared to traditional statistical methods, ML methods have greater predictive effects. ML algorithms are widely applied in nearly all steps of model establishment, such as imaging feature extraction, predictive factor classification, and model development. Therefore, this review presents the literature pertaining to ML algorithms and aims to summarize the strengths and limitations of ML, as well as its potential value in prognostic prediction, after various treatment modalities for HCC.

Background Hepatocellular carcinomas (HCCs) can be divided into proliferative and nonproliferative types, which may have implications for outcomes after conventional transarterial chemoembolization (cTACE). Biopsy to identify proliferative HCC is not routinely performed before cTACE. Purpose To develop and validate a predictive model for identifying proliferative HCCs using CT imaging features and to compare therapeutic outcomes between predicted proliferative and nonproliferative HCCs after cTACE according to this model. Materials and Methods This retrospective multicenter study included adults with HCC who underwent liver resection or cTACE between August 2013 and December 2020. A CT-based predictive model for identifying proliferative HCCs was developed and externally validated in a cohort that underwent resection. Diagnostic performance was calculated for the model. Thereafter, patients in the cTACE cohort were stratified into groups with predicted proliferative or nonproliferative HCCs according to the model. The primary outcome was overall survival (OS), and the secondary outcomes were tumor response rate and progression-free survival (PFS). These were compared between the two groups with use of the χ² test and the log-rank test. Results A total of 1194 patients (1021 men; mean age, 54 years ± 12 [SD]; median follow-up time, 29.1 months) were included. The predictive model, named the SMARS score, incorporated lobulated shape, mosaic architecture, α-fetoprotein levels, rim arterial phase hyperenhancement, and satellite lesions. The area under the receiver operating characteristic curve for the SMARS score was 0.83 for the training cohort and 0.80 for the validation cohort. According to the SMARS score, patients with predicted proliferative HCCs (n = 114) had lower tumor response rate (48% vs 71%; P < .001) and worse PFS (6.6 months vs 12.4 months; P < .001) and OS (14.4 months vs 38.7 months; P < .001) than those with nonproliferative HCCs (n = 263). Conclusion The predictive model demonstrated good performance for identifying proliferative HCCs. According to the SMARS score, patients with predicted proliferative HCCs have worse prognosis than those with predicted nonproliferative HCCs after cTACE. © RSNA, 2023 Supplemental material is available for this article.

Background: To compare the efficacy and safety between conventional transarterial chemoembolization (cTACE) and drug-eluting beads TACE (DEB-TACE) in patients with infiltrative hepatocellular carcinoma (iHCC). Methods: A total of 89 iHCC patients who were treated with either cTACE (n = 33) or DEB-TACE (n = 56) between April 2013 and September 2017 were included in this retrospective study. Patients with the situations that might have a poor outcome were defined as advanced disease including Child-Pugh class B, bilobar lesions, tumor size greater than 10 cm, ECOG 1–2, tumor burden of 50–70%, and the presence of ascites, arterioportal shunt (APS), and portal venous tumor thrombus (PVTT). The tumor response was measured 1-month and 3-month after the procedure. Progression-free survival (PFS) was calculated. Toxicity was graded by Common Terminology Criteria for Adverse Events v5.0 (CTCAE v5.0). The differences in tumor response, PFS, and toxicity were compared between the DEB-TACE group and cTACE group. Results: At 1-month and 3-month after the procedure, the objective response rate (ORR) in the overall study population was similar in DEB-TACE group and cTACE group. The disease control rate (DCR), at 1-month after the procedure, was significantly higher in the patients treated with DEB-TACE relative to those treated with cTACE (P = 0.034), while after 3 months, the difference did not differ between two groups. DEB-TACE showed a higher DCR than cTACE in patients with tumor size greater than 10 cm (P = 0.036) or associated with APS (P = 0.030) at 1-month after the procedure, while after 3 months, the difference was only noted in patients with APS (P = 0.036). The median PFS in DEB-TACE group was 96 days, while in cTACE group was 94 days, and there was no difference in PFS between two groups (P = 0.831). In the side effect analysis, abdominal pain (P = 0.034) and fever (P = 0.009) were more frequently present in the cTACE group than DEB-TACE group, but there was no difference in high grade liver toxicity between the two groups. Conclusions: Compared to cTACE, DEB-TACE offers slightly better DCR and tolerability for iHCC patients, particularly in patients associated with APS and large tumor size. However, DEB-TACE does not provide higher PFS than cTACE.

Multi-Task Learning (MTL) has attracted increasing attention in recommender systems. A crucial challenge in MTL is to learn suitable shared parameters among tasks and to avoid negative transfer of information. The most recent sparse sharing models use independent parameter masks, which only activate useful parameters for a task, to choose the useful subnet for each task. However, as all the subnets are optimized in parallel for each task independently, it is faced with the problem of conflict between parameter gradient updates (i.e, parameter conflict problem). To address this challenge, we propose a novel Contrastive Sharing Recommendation model in MTL learning (CSRec). Each task in CSRec learns from the subnet by the independent parameter mask as in sparse sharing models, but a contrastive mask is carefully designed to evaluate the contribution of the parameter to a specific task. The conflict parameter will be optimized relying more on the task which is more impacted by the parameter. Besides, we adopt an alternating training strategy in CSRec, making it possible to self-adaptively update the conflict parameters by fair competitions. We conduct extensive experiments on three real-world large scale datasets, i.e., Tencent Kandian, Ali-CCP and Census-income, showing better effectiveness of our model over state-of-the-art methods for both offline and online MTL recommendation scenarios.

Purpose: To investigate whether incomplete thermal ablation is associated with a high risk of tumor progression in patients with hepatocellular carcinoma (HCC), and to compare the efficacy of repeated thermal ablation and transarterial chemoembolization (TACE) for residual tumor after incomplete ablation.Methods: This retrospective study included 284 patients with unresectable HCC who underwent thermal ablation from June 2014 to September 2020. The response of the initially attempted ablation was classified into complete (n=236) and incomplete (n=48). The progression-free survival (PFS) and overall survival (OS) were compared between patients with complete and incomplete responses, before and after a one-to-one propensity score-matching (PSM), and between patients in whom repeated ablation or TACE was performed after a first attempt incomplete ablation.Results: After PSM of the 284 patients, 46 pairs of patients were matched. The PFS was significantly higher in the complete response group than in the incomplete response group (P<0.001). No difference in OS was noted between two groups (P=0.181). After a first attempt incomplete ablation, 29 and 19 patients underwent repeated ablation and TACE, respectively. There were no significant differences in PFS (P=0.424) and OS (P=0.178) between patients who underwent repeated ablation and TACE. In multivariate Cox regression analysis, incomplete response (P<0.001) and Child-Pugh class B (P=0.017) were independent risk factors for tumor progression, while higher AFP level (P=0.011) and Child-Pugh class B (P=0.026) were independent risk factors for poor OS.Conclusion: Although patients with incomplete ablation are associated with tumor progression compared with those with complete ablation, their OS is not affected by incomplete ablation. When patients present with residual tumors, TACE may be an alternative if repeated ablation is infeasible.