Interpretable and Open AI Models: A Mandate for the Future of HCC Diagnostics

Abstract

We read with interest the recent study by Zheng et al. [1] that introduces an innovative MRI-based topology deep learning (DL) model achieving high diagnostic accuracy in the prediction microvascular invasion (MVI), highlighting a key development in the integration of artificial intelligence (AI) into oncologic imaging. Microvascular invasion is among the critical determinants of prognosis in HCC and is typically associated with early recurrence and poor survival outcomes. Accurate prediction of MVI is therefore paramount, not only to guide therapeutic decisions [2] but also to stratify patients for clinical trials. HCC begins often as a tumour with less-invasiveness in the early stages of its development. In the stage, it grows not invasively, forming a non-cancerous tissue capsule of compressed surrounding tissue. In liver resection for HCC, resection at the capsule level is sometimes considered R0 resection. However, during the developmental process, cancer cells invade outside the capsule and into vessels, such as the portal vein. After going through this step, the aggressiveness of HCC increases rapidly, and surgical intervention based on the aforementioned recognition cannot obtain a sufficient therapeutic effect. Therefore, the macroscopic type is strongly related to its prognosis after intervention [3, 4]. It has long been pointed out the difficulty connecting these findings to surgical outcomes. Present research by Zheng et al. suggests that new insights may be obtained by connecting detailed findings (not limited to morphological changes) of MRI to topology and putting them in AI analysis. Although minute extracapsular and/or vascular invasions can be confirmed in postoperative pathology, reliable preoperative imaging modality for their early detection has not yet been established.

Although the study represents a remarkable technical achievement, it also invites broader reflection on the role and responsibilities of AI in clinical decision-making: first, the development of interpretable AI models that clinicians can trust and understand; and second, the democratisation of these tools through open-source frameworks to ensure their widespread validation and application.

The growing field of AI in medicine has transformed how we approach complex problems, particularly in diagnostic imaging. Advances in convolutional neural networks (CNNs) and, more recently, topological data analysis (TDA), have enabled models to extract nuanced patterns from imaging data, surpassing the diagnostic capabilities of traditional radiological methods [5]. In the context of HCC, the concept of a ‘virtual biopsy’ is particularly compelling. By inferring histopathological features, such as MVI from imaging data alone, these models could obviate the need for invasive procedures, reduce patient morbidity and accelerate clinical decision-making. This field is rapidly developing, and the use of a combination of clinical and quantitative imaging data has already demonstrated the great potential that a virtual biopsy can have in managing these patients [6].

However, as the complexity of AI tools increases, so can their opacity [7] for the end user. The majority of deep learning models function as ‘black boxes’, producing predictions without offering insights into their reasoning. This lack of transparency is a significant barrier to their adoption in clinical practice, where medical professionals must understand and trust the rationale behind algorithm-driven recommendations. In high-stakes decisions, such as whether to recommend anatomic hepatic resection over less invasive treatments, or to allocate patients to transplant rather than other therapies, blind reliance on machine-generated outputs is not only undesirable but also ethically irresponsible. The main question is: Are we ready to make clinical choices for patients derived from algorithms whose reasoning and conclusions we do not fully understand?

The application of TDA proposed by Zheng et al. represents a commendable step towards improving the interpretability of AI models. By encoding spatial and structural relationships within MRI data, TDA enhances the model's ability to capture biologically relevant features, bridging the gap between computational output and clinical reasoning. The incorporation of saliency maps to highlight regions contributing to MVI predictions further demonstrates a commitment to transparency. However, the achieved interpretability still remains insufficient unless paired with broader accessibility and clinical integration.

Moreover, the utility of these models is often constrained by their proprietary nature. Closed-source algorithms hinder independent validation and limit their applicability to broader patient populations.

In the case of the model presented in this publication, its training and validation were restricted to patients within BCLC stages A and B, a subgroup that does not encompass the full spectrum of HCC presentations.

This is an important limit, as the presence of MVI can be detected in any stage. According to the Italian national register (HE.RC.O.LE.S.), it has been detected in 58.2% among BCLC 0-A, 47.9% among BCLC B and 41.9% among BCLC-C cases (Table 1). Furthermore, the BCLC therapeutic decision algorithm has recently been totally superseded by the ‘therapeutic hierarchy’ approach [8], which clearly indicates that surgery can be the first-line treatment regardless of the oncological stage. In this setting, limiting the possibility of applying the virtual biopsy proposed by the authors only to cases with less advanced disease, severely limits the applicability of the algorithm itself in everyday practice. The limitation of this model to BCLC stages A and B patients exemplifies a recurring limitation in AI-driven research: the disconnection between experimental design and real-world data diversity. Clinicians treating HCC encounter a spectrum of presentations, from early-stage lesions amenable to curative resection to advanced, multifocal disease requiring palliative interventions. Models confined to narrow patient subsets risk perpetuating inequities in access to care, particularly for underserved populations disproportionately affected by advanced-stage HCC.

Another significant limitation lies in the study's proprietary nature. To fully realise the promise of AI in medicine, the development of open, interpretable and universally accessible models is non-negotiable. Although the authors report promising external validation results, independent replication by other institutions is essential to establish robustness and generalisability and should be encouraged at best. This result cannot be achieved simply with data from two centres, but it is necessary to freely allow the use of the model. Although open-source frameworks foster collaboration and enhance the robustness of AI models, the development and deployment of such tools must be guided by clear regulatory frameworks. Without structured oversight, the integration of diverse and potentially biased datasets into open models risks introducing variability that may confuse real-world clinical applications. Establishing guidelines to manage data heterogeneity and ensure model reliability across populations is essential for mitigating these challenges.

In conclusion, the study by Zheng et al. underscores the transformative potential of AI in HCC management but also highlights the ethical and practical challenges that accompany these advancements. Interpretability and openness are not ancillary concerns; they are essential to the responsible integration of AI into medicine. As we navigate the rapidly evolving landscape of AI-driven diagnostics, we must remain steadfast in our commitment to these principles, ensuring that technological progress reliably serves the human elements of care.

The authors declare no conflicts of interest.