Liver International, Journal Year: 2025, Volume and Issue: 45(4)
Published: March 14, 2025
Early-stage hepatocellular carcinoma (HCC), defined as a single tumour or up to three lesions < 3 cm, according the Barcelona Clinic Liver Cancer (BCLC) classification, is eligible for curative treatment [1]. Curative options include liver transplantation, surgical resection percutaneous ablation. Ablation can be selected first-line therapy over surgery tumour(s) cm due its advantages, such lower complication rates, reduced mortality and minimal invasiveness, making it feasible even when function slightly impaired in presence of portal hypertension. Moreover, transplant-eligible patients, strategy ablation followed by salvage transplantation case recurrence gaining traction worldwide context graft shortages [2]. However, unlike surgery, does not allow full histopathological analysis resected specimen, limiting tissue samples obtained via micro-biopsies prior This limitation prognostic stratification following first treatment, especially transplantable recognised histological markers aggressiveness, microvascular invasion satellites, cannot captured biopsies. Intratumoral heterogeneity (ITH) another promising marker with potential provide valuable information on prognosis risk early recurrence. As ITH only fully assessed through genomic evaluation entire lesion, imaging features ITH, using modalities radiomics and/or deep-learning, could serve surrogate advanced HCC early-stage HCCs treated where, definition, no specimen available. In this issue International, Zhang et al. describe transformer-based quantitative model that integrates signatures extracted from ultrasound (US), contrast-enhanced US (CEUS) magnetic resonance (MRI), acquired before ablation, along demographic, clinicopathological laboratory variables predict individual [3]. The was tested cohorts patients radiofrequency (RFA) microwave (MWA), then validated external undergoing RFA, laser (LA) irreversible electroporation (IRE) treatments. primary structure network used extract ITH-related referred vision-transformer-based intratumoral (ViT-Q-ITH) model. deep learning segments images into patches employs self-attention mechanism analyse correlations between these patches, capturing both global local relationships enhance understanding complex visual structures. A combined developed integrating ViT-Q-ITH score clinical factors. study results show achieved high performance internal validation cohort (AUC 0.86) test 0.83), sensitivities 76% 74% specificities 88% 84%, respectively, outperforming traditional standalone Recurrence-free survival analyses further confirmed superior capability model, clearly distinguishing high-risk low-risk more accurately than models alone. all cohorts, identified had significantly higher probability (local distant) compared patients. One study's strengths lies generalisability applied 'real-world' data cohorts. Indeed, differed therapeutic employed (some IRE LA instead RFA/MWA) but also baseline biological characteristics (e.g., much smaller tumours less disease) training cohort. Despite unfavourable conditions paper, showed excellent performance, achieving AUCs 0.8. suggests may applicable across different settings modalities, broadening use. An innovative aspect development data, known linked recurrence, readily available pictures CEUS. It worth noting two MRI sequences were unenhanced acquisitions (T2- diffusion-weighted imaging), opposed numerous studies imaging, which generally vascularity/texture volumes. choice likely made limit variations related image contrast, injection type timing, thereby improving model's centres. one recurrent challenge widespread adoption AI-based their interpretability. Although vision-transformer offers neural nature makes interpret approaches based simple static images. critical because predictive must understandable AI specialists clinicians who want use them daily practice. challenges, appears robust, there are several areas improve applicability, particularly extending multi-institutional international datasets robustness diverse populations. target population biomarkers order stratify, predicted risk, whether they should fast-tracked rather an 'ablate wait' strategy. Additionally, other sources, genetic sequencing circulating biomarkers, power. Combining complete picture biology, prediction accuracy enabling personalised stratification. Lastly, interpretability addressed thoroughly. Clinician involvement help develop transparent tools practice, reducing common feeling among physicians dealing 'black box' tools. context, explainable (XAI) techniques will play crucial role, maintain without sacrificing mechanisms [4]. conclusion, contributes growing body evidence supporting [5]. While challenges remain, terms large-scale validation, have clinicians' ability tailor patient management accordingly. authors declare conflicts interest. Data sharing article, new created analyzed study.
Language: Английский