Vision transformer-based model can optimize curative-intent treatment for patients with recurrent hepatocellular carcinoma

Abstract

The treatment selection for recurrent hepatocellular carcinoma (rHCC) within Milan criteria after hepatectomy remains challenging. Here, we present HEROVision, a Vision Transformer-based model designed for personalized prognosis prediction and treatment optimization between thermal ablation (TA) and surgical resection (SR). HEROVision is trained on initial HCC cohorts (8492 images; 772 patients) and independently tested on rHCC cohorts (9163 images; 833 patients) from five centers. Propensity score matching (PSM) forms two groups of rHCC patients underwent TA and SR to fairly evaluate whether optimized treatment selection by HEROVision have clinical benefits. HEROVision significantly outperforms all six guideline staging systems in the external testing cohort, both in time-dependent concordance index and area under the curve (all P < 0.002). After PSM, 35.9% (23/64) and 6.6% (6/91) high-risk rHCC patients are identified, who could achieve improved prognosis by changing their treatments. HEROVision shows promise in optimizing individualized treatment between TA and SR for early-stage rHCC, complementing current clinical guidelines.

Fig. 1. Flowchart of patient inclusion and exclusion.

iHCC initial hepatocellular carcinoma, rHCC recurrent hepatocellular carcinoma, TA thermal ablation, SR surgical resection.

Fig. 2. Workflow of development and clinical application of HEROVision.

a A real-world clinical decision-making scenario for rHCC. b Construction of HEROVision models for TA and SR using the 1stTA and 1stSR datasets, respectively. c Clinical application of HEROVision-TA and -SR models. Cross-stratifications were executed by employing HEROVision-TA and -SR models on CEUS and MRI from exchanged patient cohorts, respectively. Subsequently, individualized prediction of PFS was compared between the original and re-stratified patient subgroups for each of the 2ndTA and 2ndSR datasets. iHCC initial hepatocellular carcinoma, rHCC recurrent hepatocellular carcinoma, CEUS contrast-enhanced ultrasound, MRI magnetic resonance imaging, TA thermal ablation, RHR repeat hepatic resection, SR surgical resection, CAD computer-aided diagnosis, T2WI T2-weighted imaging, DWI diffusion-weighted imaging, PFS progression-free survival, HEROVision hepatocellular carcinoma optimizing treatments with Vision Transformer network.

Fig. 3. Kaplan–Meier analysis of PFS stratified by the HEROVision-TA and -SR models into low- and high-risk groups.

a–c TA dataset in the training, internal validation, and external testing cohorts, respectively. d–f SR dataset across all three cohorts. The error bands (dashed lines) represent the 95% confidence intervals, and the P values were calculated using the two-sided Log-rank test. Source data are provided as a Source Data file. TA thermal ablation, SR surgical resection, PFS progression-free survival, HR hazard ratio.

Fig. 4. Comparisons between HEROVision and six major guideline methods.

For TA (a, b) and SR (c, d) datasets, the C-index values in the training, internal validation, and external testing cohorts (a, c), along with the time-dependent AUCs in all three cohorts (b, d), shows that HEROVision consistently outperformed all six major staging systems proposed by different guidelines for prognostic prediction. Furthermore, there were noticeable enhancements in 2-year C-index and AUC for the six staging systems, after they were supplemented with HEROVision for both 2ndTA (e, f) and 2ndSR (g, h) datasets. Source data are provided as a Source Data file. TA thermal ablation, SR surgical resection, C-index concordance index, AUC area under the curve, AJCC American Joint Committee on Cancer, BCLC Barcelona Clinic Liver Cancer, CNLC China Liver Cancer, HKLC Hong Kong Liver Cancer, ITA.LI.CA Italian Liver Cancer, UICC Union for International Cancer Control.

Fig. 5. Optimizing treatment selection between TA and RHR for rHCC by HEROVision.

a In the matched 2ndTA group, 23 out of 214 rHCC patients changed their risk categories after re-stratification by HEROVision-SR. Among them, 23 original high-risk patients could be downgraded to low-risk if they switched from TA to RHR (displayed by the orange to blue branch). b Comparison of predicted risk scores between the 64 high-risk patients and the 23 of them who required a change of treatment in the 2ndTA group. c Kaplan–Meier curves of PRS stratified by HEROVision-TA in the matched 2ndTA group. d The same analysis as (c) in the matched 2ndTA group after removing the 23 identified patients. e In the matched 2ndSR group, 10 out of 214 rHCC patients changed their risk categories after re-stratification by HEROVision-TA. Among them, six original high-risk patients could be downgraded to low-risk if they switched from RHR to TA (displayed by the orange to blue branch). f Comparison of predicted risk scores between the 91 high-risk patients and the six of them who required a change of treatment in the 2ndSR group. g Kaplan–Meier curves of PRS stratified by HEROVision-SR in the matched 2ndSR group. h The same analysis as (g) in the matched 2ndSR group after removing the six identified patients. P values were computed using the two-sided Mann–Whitney U test (b, f). Boxes indicate the upper and lower quartiles (Q3 and Q1), with a line at the median. Whiskers extend to the maximum and minimum values within 1.5 times the interquartile range. Outliers are shown as circles and identified via the interquartile range rule. The error bands (dashed lines) represent the 95% confidence intervals, and the P values were calculated using the two-sided Log-rank test (c, d, g, h). Source data are provided as a Source Data file. TA thermal ablation, SR surgical resection, RHR repeat hepatic resection, PFS progression-free survival, PRS post-recurrence survival.