AI and Cancer Research ── Liver ── 2026-10-25
Liver cancer: treatment, trials and AI
- Abstract
Rationale and objectives This study aims to develop a framework integrating dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI)-based multi-channel deep learning combined with clinical modeling for the non-invasive preoperative discrimination between hepatocellular carcinoma (HCC) and mass-forming intrahepatic cholangiocarcinoma (MF-ICC).
Methods: The internal development cohort comprised 334 surgically confirmed patients, including 231 patients with HCC and 103 with MF-ICC, and was divided at the patient level into a training cohort ( n = 233) and an internal validation cohort ( n = 101). The independent external validation cohort included 50 patients, comprising 40 with HCC and 10 with MF-ICC. We constructed separate two-dimensional deep learning models based on arterial phase, portal venous phase, and delayed phase images and then developed fusion models (2D model). A multi-channel deep learning model was further developed by directly inputting the three-phase images stacked along the channel dimension. Moreover, a clinical model was built based on independent predictors identified through multivariate analysis, which was ultimately integrated with the best-performing deep learning model. Model performance was evaluated and compared.
Results: Deep learning model directly feeding the three-phase images as a multi-channel input achieved better discrimination (area under the curve [AUC] = 0.814 in the internal validation cohort) than all single-phase and corresponding fusion models. The comprehensive 2D + multi-channel + clinical model showed the best performance for the differentiation of HCC from MF-ICC, achieving an AUC of 0.887 (95% CI, 0.819–0.954) in the internal validation cohort, and generalized well to the external validation cohort (AUC = 0.908; 95% CI, 0.811–1.000). However, no significant differences in AUCs were found among the models after correction for multiple comparisons.
Conclusion: The fusion model integrating clinical predictors with DCE-MRI multi-phase deep learning feature may provide a valuable tool for preoperatively differentiating HCC from MF-ICC.
Journal IF-equivalent: 2.7 (OpenAlex 2-year mean citedness, value as of 2026-10-02, retrieved 2026-10-03; not the official Clarivate IF)Reference: Zhang H, Guo L, Jia F, Liu W, Zhang M, Li F, et al. Dynamic contrast-enhanced MRI deep learning model combined with clinical model for noninvasive differentiation of hepatocellular carcinoma from mass-forming intrahepatic cholangiocarcinoma. Front Oncol. 2026 Oct 23;16:1936989. doi:10.3389/fonc.2026.1936989.Checked: Abstract only - Abstract
Introduction: Clinical responses to atezolizumab plus bevacizumab (AB) therapy can vary across hepatocellular carcinoma (HCC) nodules among individual patients. This study aimed to derive and externally validate a non-invasive radiomics signature (ABRad) translated from a transcriptome signature for nodule-level prediction of AB response (ABS).
Methods: We originally derived and validated ABS using 209 and 44 pre-treatment HCC samples from the IMbrave150 and GO30140 trials, respectively. ABS was then translated into ABRad using a machine-learning approach on 59 patients with paired genome-wide transcriptome and pre-resection computed tomography imaging from surgically resected HCC. Subsequently, ABRad was externally validated using 178 HCC nodules from 55 patients treated with AB therapy.
Results: The 148-gene ABS predicted therapeutic response and progression-free survival (PFS) in the clinical trial cohorts. The final ABRad model, constructed using a random forest algorithm, was heavily weighted toward contrast-enhanced features. In the external validation cohort, ABRad demonstrated a significant positive association with the degree of therapeutic response (p <0.001). The areas under receiver operating characteristics curve for predicting complete response (CR), objective response (OR), and disease control (DC) were 0.71 (95% confidence interval [CI], 0.61-0.80), 0.65 (95% CI, 0.56-0.73), and 0.69 (95% CI, 0.59-0.78), respectively. Nodule-level ABRad was independently associated with therapeutic response with adjusted odds ratios of 3.05 (95% CI, 1.23-7.57) for CR, 2.24 (95% CI, 1.06-4.77) for OR, and 2.79 (95% CI, 1.07-7.29) for DC. Patient-level ABRad was significantly associated with prolonged PFS (p = 0.005).
Conclusions: ABRad successfully predicted nodule-level response to AB therapy in HCC.
Journal IF: unknown (could not be matched)Reference: [No authors listed]. Supplemental Material for: A radio-transcriptome signature for nodule-level prediction of response to atezolizumab plus bevacizumab in hepatocellular carcinoma. Figshare. [Epub ahead of print]. doi:10.6084/m9.figshare.34754340.Checked: Abstract only - Abstract
Objectives: To develop a radiomics‑based predictive model for posthepatectomy liver failure (PHLF) using liver shear-wave elastography (SWE) and to evaluate its performance against conventional clinical indicators.
Methods: This retrospective study enrolled 948 patients with pathologically confirmed hepatocellular carcinoma who underwent perioperative laboratory tests and SWE at the Third Affiliated Hospital of Naval Medical University (Eastern Hepatobiliary Surgery Hospital) between December 2021 and March 2025. Patients were randomly divided into training (n = 664) and validation (n = 284) cohorts (7:3 ratio). Radiomics features were extracted from SWE images of tumor and adjacent paratumor regions. Four models—tumor rad‑signature, paratumor rad‑signature, clinical, and conjunctive (integrating radiomics and clinical features)—were constructed using these features alongside clinical variables. Multiple machine learning algorithms were applied, and all model selection, including algorithm comparison and hyperparameter tuning, was performed exclusively via cross‑validation on the training cohort; the validation cohort was used only for independent performance evaluation.
Results: The locked tumor rad-signature and paratumor rad-signature models yielded area under the curve values of 0.846 and 0.824, respectively, in the independent validation cohort. Multivariate logistic regression identified total protein, prealbumin, prothrombin time, paratumor elasticity value, and transfusion as independent predictors of PHLF. The conjunctive model achieved area under the curve of 0.939 in the training cohort and 0.886 in the independent validation cohort. Calibration curves showed good agreement between predicted and observed probabilities, and a decision curve analysis suggested a potential clinical net benefit in both cohorts.
Conclusions: Radiomics models based on liver SWE demonstrated promising predictive ability for PHLF, suggesting a potential role of tissue elasticity in risk stratification. Although this approach may offer value for perioperative risk assessment and preliminary decision support for perioperative management, the analyses should be considered exploratory, and further external validation is necessary before any clinical implementation.
Journal IF-equivalent: 0.3 (OpenAlex 2-year mean citedness, value as of 2026-10-23, retrieved 2026-10-24; not the official Clarivate IF)Reference: Gu Y, Ren JL, Qian Y, Chen Y, Xu J, Li Y, et al. Ultrasound Elastography Radiomics as a Predictor of Posthepatectomy Liver Failure. Hemodynamics. 2026 Oct 22 [Epub ahead of print]. doi:10.1097/hd9.0000000000000037.Checked: Abstract only - Abstract
Noninvasive preoperative prediction of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) remains a significant clinical challenge worldwide. This study encoded feature variations across multi-phase MRI sequences for the preoperative prediction of MVI. Data were retrospectively examined for 115 consecutive patients (mean age: 57.2 ± 10.7 years; 95 men) who had undergone surgical resection for HCC (tumor diameter less than 3 cm). All of them underwent preoperative MR examination and were randomly partitioned into training and testing cohorts at an 8:2 ratio. Peritumoral expansions (3 mm, 5 mm, and 10 mm) were generated from manually segmented intratumoral regions of interest (ROIs) across the arterial, venous, and delayed phases. Radiomics features were subsequently extracted from both the original and expanded ROIs. We then evaluated the predictive performance of five encoding methods combined with four machine learning classifiers using both clinical characteristics and radiomics features. Multi-phase models incorporating temporal encoding methods consistently outperformed single-phase models and those relying solely on clinical characteristics. Among all evaluated combinations, the Fourier-inspired temporal encoding method combined with an ExtraTrees classifier, applied to the original tumoral ROI with a 5-mm peritumoral expansion, achieved the best diagnostic performance, with an AUC of 0.850 (95% CI: 0.675–0.978). While radiologists routinely monitor multi-phase contrast dynamics qualitatively, our findings demonstrate that these cross-phase variations provide quantifiable, highly predictive information for MVI. This framework represents a promising, noninvasive quantitative tool for preoperative MVI risk stratification.
Journal IF-equivalent: 3.8 (OpenAlex 2-year mean citedness, value as of 2026-10-02, retrieved 2026-10-03; not the official Clarivate IF)Reference: Dou Y, Liu Y, Yuan W, Wu S, Yao Z. Preoperative prediction of microvascular invasion in small hepatocellular carcinoma (≤ 3 cm) by leveraging cross-phase variations of MRI contrast phases. BMC Med Imaging. 2026 Oct 22 [Epub ahead of print]. doi:10.1186/s12880-026-02984-y.Checked: Abstract only