AI Imaging Meets Flow Cytometry in Race to Detect Liver Cancer Cells in Blood

A team of researchers at the University of Fukui in Japan has put two very different ways of hunting cancer cells in the bloodstream head to head, and the results suggest that the future of the liquid biopsy may belong to neither technology alone. In a study published in BMC Cancer, Juan Lyu, Takuto Nosaka, Yasunari Nakamoto and colleagues evaluated, in parallel, an artificial intelligence–assisted fluorescence imaging platform known as TruAI and conventional fluorescence-activated cell sorting, or FACS, for the detection of circulating tumor cells in patients with hepatocellular carcinoma, the most common form of primary liver cancer and one of the deadliest malignancies worldwide. Their central finding is striking in its simplicity: the two platforms count cells with remarkable agreement, but only the imaging approach can see what the cells actually look like.

Circulating tumor cells are cancer cells that shed from a primary tumor, survive the hostile currents of the bloodstream and can eventually seed new metastases. Because they can be captured from a simple blood draw, they have long been considered one of the most promising liquid biopsy biomarkers in oncology. In hepatocellular carcinoma, where tumors often arise in livers damaged by hepatitis B, hepatitis C or other chronic disease and where tissue biopsies can be risky, a blood test that reliably tracks tumor burden would be transformative. But circulating tumor cells are vanishingly rare, often fewer than a handful per milliliter of blood amid billions of ordinary blood cells, and detecting them demands both extreme enrichment and extremely sensitive identification.

The Fukui team’s workflow began with CD45 depletion–based enrichment, a technique that removes the white blood cells marked by the CD45 surface protein and leaves behind a concentrate enriched for rare non-immune cells, including tumor cells. Both platforms then interrogated this enriched fraction using a common immunophenotypic definition: a cell was counted as a circulating tumor cell if it carried DAPI nuclear staining, indicating intact DNA within a nucleus; lacked CD45, confirming it was not a leukocyte; and expressed pan-cytokeratin, an epithelial marker that liver cancer cells retain even after they leave the tumor. The requirement of preserved nuclear morphology added an important quality gate, filtering out dying cells and debris that might otherwise inflate counts.

FACS, the workhorse of modern cell analysis, pushes thousands of cells per second through a laser interrogation point, measuring fluorescence on multiple channels simultaneously. Its strengths are throughput, multiparametric precision and the ability to quantify absolute cell numbers with statistical rigor. Its weakness is blindness: it reduces each cell to a vector of intensity values, discarding the spatial and structural information that pathologists have relied on for more than a century. TruAI, an image analysis platform built on deep learning, approaches the same enriched sample from the opposite direction. It scans fields of view under a fluorescence microscope and a trained neural network classifies candidate cells, weighing nuclear shape, staining patterns and cytokeratin architecture in a way that mimics, and in principle exceeds, human visual inspection.

To validate that both platforms could actually recover known numbers of tumor cells, the researchers performed spike-in experiments, mixing defined quantities of cultured hepatocellular carcinoma cell lines into blood samples before enrichment. Both methods returned counts that tracked the spiked concentrations with strong linearity, with coefficients of determination exceeding 0.94, meaning that more than 94 percent of the variance in measured counts was explained by the true number of cells present. Notably, FACS consistently registered higher absolute counts than TruAI across the dilution series, a discrepancy the authors interpret in light of the platforms’ different gating strategies and sensitivity thresholds, and one that underscores why head-to-head calibration matters when technologies are used interchangeably in clinical studies.

The clinical core of the study involved 17 patients with hepatocellular carcinoma whose peripheral blood was sampled and analyzed on both platforms. When the counts were compared, the concordance was unmistakable: TruAI-derived and FACS-derived circulating tumor cell counts correlated significantly, with a Pearson correlation coefficient of 0.8216 and an R-squared of 0.675, significant at p below 0.0001. In practical terms, when a patient’s blood registered high on one platform, it registered high on the other, and when treatment drove counts down on one, the other followed. This is precisely the kind of analytic agreement that regulatory scientists look for when deciding whether two assays measure the same biological quantity, and it suggests that either platform could serve as a quantitative readout of circulating tumor burden.

But the study did not stop at static counting. Several of the patients were receiving systemic therapy, including the atezolizumab plus bevacizumab combination that has become the standard first-line regimen for advanced hepatocellular carcinoma, and the team tracked their circulating tumor cells longitudinally across treatment. Both platforms captured the same direction of change, documenting significant decreases in circulating tumor cell numbers as therapy took hold. The convergence of two independently derived measurements on the same biological trend strengthens confidence that the decline was real rather than a processing artifact, and it demonstrates that serial blood-based monitoring is technically feasible with either approach.

The decisive advantage, however, belonged to TruAI, and it was morphological. Because the imaging platform retains the architecture of each detected cell, the researchers could distinguish solitary circulating tumor cells from circulating tumor cell clusters, small aggregates of two or more tumor cells traveling together. Clusters are widely regarded as the more dangerous currency of metastasis; experimental work across multiple tumor types has shown that cell clusters seed metastases far more efficiently than single cells, and cluster burden has been associated with worse prognosis. In the Fukui cohort, the proportion of circulating tumor cells found in clusters fell in patients who achieved disease control, and those reductions in cluster proportion ran in parallel with radiologically confirmed tumor shrinkage and with declining levels of serum tumor markers, including alpha-fetoprotein and des-gamma-carboxy prothrombin, the two canonical biomarkers of hepatocellular carcinoma activity.

That parallelism is the quiet headline of the study. A blood test that reports not just how many tumor cells are circulating but how they are organized, and that tracks those organizational changes in step with what radiologists see on CT and MRI, moves the liquid biopsy closer to functioning as a genuine surrogate of treatment response. For patients with hepatocellular carcinoma, where response assessment currently depends on imaging performed at fixed intervals and where serum markers can be insensitive in a substantial fraction of cases, a morphologically informative blood assay could eventually allow earlier detection of treatment failure and faster escalation to second-line options. The authors are careful about scope: their conclusions position the combined use of TruAI imaging and flow cytometry as a way to enhance longitudinal monitoring by capturing both total circulating tumor cell burden and cluster dynamics, not as a validated diagnostic or prognostic test ready for routine deployment.

The study’s limitations are those of an early-phase evaluation: a small cohort of 17 paired clinical samples, a single institution, and a sensitivity analysis using all 31 paired samples reported in supplementary material rather than a multicenter validation. Yet the technical message is already clear and, in a field crowded with competing liquid biopsy platforms, unusually constructive. Counting and seeing are not rival philosophies but complementary layers of the same measurement. Flow cytometry supplies speed, absolute quantification and multiparametric depth; AI-driven imaging supplies structure, morphology and the ability to recognize biologically meaningful subpopulations such as clusters that a flow cytometer sweeps past without notice. As deep learning image analysis matures and regulatory pathways for AI-based diagnostics take shape, the Fukui results argue that the most powerful blood test for liver cancer may be the one that learns to do both at once.

Subject of Research: AI-assisted imaging and flow cytometry for circulating tumor cell detection in hepatocellular carcinoma

Article Title: Parallel Evaluation of TruAI-assisted imaging and flow cytometry for circulating tumor cell detection in hepatocellular carcinoma

Article References: Lyu, J., Nosaka, T., Nomiya, H., Murata, Y., Akazawa, Y., Tanaka, T., Takahashi, K., Naito, T., Ohtani, M., Zhang, L., & Nakamoto, Y. (2026). Parallel Evaluation of TruAI-assisted imaging and flow cytometry for circulating tumor cell detection in hepatocellular carcinoma. BMC Cancer. https://doi.org/10.1186/s12885-026-17005-y

Image Credits: AI Generated

DOI: 10.1186/s12885-026-17005-y

Keywords: hepatocellular carcinoma, circulating tumor cells, liquid biopsy, TruAI, flow cytometry, artificial intelligence, deep learning, CTC clusters, atezolizumab, bevacizumab, biomarkers, BMC Cancer