The study focused on a demanding clinical setting: CT-guided renal cryoablation, a minimally invasive procedure in which physicians thread thin, needle-like probes into kidney tumors to freeze and destroy cancerous tissue. These procedures require repeated CT imaging—first to plan the approach, then to confirm probe placement, then to monitor the freezing process and verify the result. Because patients with kidney cancer often undergo serial imaging for diagnosis, treatment, and years of follow-up, their cumulative radiation exposure can add up quickly. Reducing the dose of each scan, even modestly, compounds into a meaningful protective benefit over a patient’s entire course of care.

The researchers designed an elegant experiment that sidestepped a common pitfall in dose-reduction research. Traditionally, scientists either simulate low-dose scans by injecting mathematical noise into routine-dose data, or they irradiate volunteers twice at different dose levels to compare the results directly. The first approach is limited by restricted access to raw projection data; the second raises ethical concerns about unnecessary exposure. Instead, the Mayo team exploited a natural feature of the ablation workflow itself. Each patient already receives a routine-dose planning scan and a lower-dose probe placement check within the same hour, on the same scanner, with the same body habitus. By retrospectively collecting both datasets from fifteen patients, the investigators achieved the statistical power of a paired, dual-acquisition design without adding a single extra milligray of radiation.

The numbers behind the comparison are straightforward. The routine-dose planning scans delivered a volumetric CT dose index of 18.9 plus or minus 4.7 milligrays, while the low-dose probe placement scans delivered 9.9 plus or minus 2.5 milligrays—roughly a fifty percent reduction. The routine-dose images were reconstructed with iDose, a Philips iterative reconstruction algorithm that represents the current clinical standard. The low-dose images were processed with Spectral Precise Image, a newly developed AI-based spectral reconstruction technique that this study was the first in the world to evaluate. Both were benchmarked against filtered back projection, the classical reconstruction mathematics that serves as a common reference point for measuring how much noise each modern algorithm removes.

The quantitative results were striking. Measured in the aorta at the level of the twelfth thoracic vertebra, iterative reconstruction reduced image noise by about 23 percent relative to filtered back projection, at both dose levels. The AI-based spectral reconstruction cut noise by roughly 61 to 62 percent—nearly three times the noise suppression of the clinical standard. Most remarkably, the low-dose images processed by AI were actually 32 percent less noisy than the routine-dose images processed with iterative reconstruction, a difference that was highly statistically significant. In other words, the AI algorithm did not merely compensate for the halved dose; it produced a cleaner image than the higher-dose scan.

Noise reduction alone means little if it comes at the cost of blurred anatomy, a known hazard of aggressive denoising that can give tissue a plastic, unnatural texture. To check for this, the researchers plotted line profiles—mathematical cross-sections of signal intensity—across a renal lesion and through the vertebral foramen, testing how well each algorithm preserved edges in both soft tissue and bone. Compared with filtered back projection, neither iterative nor AI-based reconstruction showed any discernible degradation of anatomic structures. Edge sharpness was well preserved for low-contrast and high-contrast features alike, suggesting the AI engine suppresses noise without smearing the fine boundaries radiologists depend on to localize tumors and plan probe trajectories.

The human verdict came from a blinded reader study. Two radiologists independently rated overall image quality for conventional, 50 kiloelectron volt virtual monoenergetic, and electron density results on a five-point Likert scale, without knowing which dose level or reconstruction algorithm had produced each image. For the conventional images, the low-dose AI protocol scored significantly higher than the routine-dose iterative protocol, at 3.6 versus 3.2, a difference the authors described as superior image quality for the reduced-dose condition. For the 50 keV monoenergetic images, the two protocols were statistically indistinguishable, with both scoring near 3.8 to 3.9. Only the electron density maps, which are intended for quantitative measurement rather than direct visual reading, scored low across the board, with radiologists noting that their texture appeared less realistic.

The technical achievement rests on how the AI reconstruction differs from its predecessors. Spectral CT, also called dual-energy CT, acquires x-ray data at high and low energies, enabling the generation of virtual monoenergetic images, electron density maps, and material-specific images such as iodine or calcium maps. But splitting the data by energy amplifies noise, because the increased dimensionality of energy-selective imaging inherently increases signal variance. Earlier AI reconstruction methods tackled noise in the image domain with convolutional neural networks, or in the projection data before filtered back projection, or by converting sinograms directly into images with deep learning. The Spectral Precise Image technique evaluated here is distinctive in that a single AI model can be applied across all available spectral results—conventional, monoenergetic, and electron density—providing uniform noise control across the entire spectral output of the scanner.

The implications extend well beyond kidney ablations. Interventional radiology is a particularly fertile ground for dose reduction because procedures involve many sequential scans, exposing not only patients but also physicians and care teams to scattered radiation. The authors emphasize that their study establishes feasibility rather than clinical readiness: the cohort of fifteen patients at a single center was small, the patient sizes ranged only from about 29 to 38 centimeters in water equivalent diameter, and the study did not assess diagnostic quality, in which the existence and location of lesions are unknown rather than already established. Slices containing the ablation probes were also excluded to keep readers blinded, a necessary compromise the authors acknowledge as a limitation. Future work, they write, will examine whether diagnostic CT can safely adopt the same dose reductions, how performance varies across body habitus, and how much the AI approach outperforms iterative reconstruction when doses are matched.

Still, the central result is hard to ignore: an AI algorithm, applied to scans using half the radiation, produced images that radiologists judged better than the clinical standard. As deep learning reconstruction matures from a promising laboratory concept into a validated clinical tool, the long-standing assumption that image quality must be purchased with radiation dose is beginning to crumble. For the millions of patients who undergo CT guidance each year—and for the cancer survivors who face decades of surveillance imaging—that crumbling cannot come soon enough.

Subject of Research: AI-based spectral CT reconstruction for radiation dose reduction in image-guided renal ablation

Article Title: Feasibility of radiation dose reduction with AI-based spectral reconstruction for computed tomography

Article References: Feasibility of radiation dose reduction with AI-based spectral reconstruction for computed tomography. (n.d.). https://doi.org/10.1007/s44343-025-00021-3

Image Credits: AI Generated

DOI: 10.1007/s44343-025-00021-3

Keywords: computed tomography, AI reconstruction, radiation dose reduction, spectral CT, deep learning, interventional radiology, renal cryoablation, image quality, virtual monoenergetic imaging, denoising, Mayo Clinic, medical imaging

The core insight behind DePass is deceptively simple: instead of treating noise removal and data integration as two separate problems to be solved in sequence, the framework couples them into a single iterative loop. Raw omics data are first organized into graphs, mathematical structures in which each cell or spatial spot is a node and edges connect measurements that are likely to be biologically related. The system then performs neighbourhood aggregation, borrowing information from connected nodes to smooth out spurious fluctuations in each modality. Crucially, the denoised data feed into an integration module that produces a unified embedding, a compact mathematical representation of each cell that fuses information across all measured layers. That embedding, in turn, is used to refine the enhancement step, and the cycle repeats. Each pass through the loop sharpens both the cleaned data and the integrated result, which is why the authors describe their architecture as dual-enhanced.

This coupled design addresses a weakness that has plagued earlier integration methods. Most existing tools were built with the assumption that the input data are reasonably clean, and they focus their mathematical machinery on aligning modalities that measure different aspects of biology. Yet multimodal experiments are inherently noisier than single-modality profiling, because each additional measurement layer introduces its own sources of dropout, batch effects and stochastic variation. When a framework ignores this elevated noise, the errors propagate into the integrated embedding, blurring the very biological distinctions the analysis is meant to reveal. By iteratively denoising while integrating, DePass prevents that degradation and, according to the authors’ benchmarks, recovers sharper and more faithful cellular structure.

The second major innovation is scope. Many current integration methods are tailored to a specific data type: some excel at single-cell experiments, where each cell carries paired measurements across modalities, while others are designed for spatial platforms, where molecular readouts are mapped onto tissue coordinates. DePass was engineered from the ground up to handle both contexts within a single framework. Whether the input is a single-cell multi-omics dataset, a spatial transcriptomics and proteomics experiment, or a spatial epigenomic profile, the same graph learning architecture applies. This generality matters for a field that is expanding explosively, with new experimental platforms appearing every year and laboratories increasingly unwilling to learn a different computational tool for each one.

To substantiate claims of generality, the team mounted an unusually comprehensive benchmarking campaign. DePass was evaluated across six modalities, nine tissue types and thirteen experimental platforms, spanning datasets from mouse brain and spleen to human tonsil, mouse embryo and human tumour tissue. The platforms ranged from established commercial systems such as 10x Genomics multiome assays to cutting-edge methods including Spatial-mux-seq, which captures four modalities simultaneously in mouse embryo tissue, and MISAR-seq, which jointly profiles gene expression and chromatin accessibility in spatial context. Across this gauntlet of tests, DePass demonstrated superior integration accuracy compared with competing methods, measured by quantitative metrics of how well cells cluster by biological identity rather than technical artifact.

Among the most striking demonstrations is the framework’s handling of spatial metabolomics. The researchers applied DePass to a mouse brain dataset combining Visium spatial transcriptomics with MALDI mass spectrometry imaging, a pairing that captures both the transcriptomic and metabolic state of tissue in register. Metabolomic data are notoriously noisy and sparse, making them a stern test of any denoising approach. DePass not only integrated the two layers coherently but also strengthened the correlations between genes and metabolites, sharpening anatomical boundaries in the enhanced data compared with the raw measurements. Similar gains appeared in a mouse spleen spatial transcriptomics and proteomics dataset, where enhancement improved the spatial coherence of marker gene and protein expression patterns.

The framework also proved its worth on clinical material. The team generated an in-house colorectal cancer dataset using Stereo-CITE-seq, a technology that profiles gene expression and surface proteins across intact tumour tissue at high resolution. When DePass was applied to this data, it uncovered substructure within immune niches that had been invisible in less carefully integrated analyses, and it resolved spatial heterogeneity within the tumour at near single-cell resolution. For cancer researchers, this kind of resolution is not a luxury. Tumours are mosaics of malignant cells, immune infiltrates, fibroblasts and vasculature, and the arrangement of these components shapes how the disease progresses and how patients respond to therapy. A tool that can faithfully integrate transcriptomic and proteomic views of that architecture could sharpen the identification of biomarkers and therapeutic targets.

Further applications underscored the breadth of the method. In hepatocellular carcinoma data combining Xenium transcriptomics with CODEX protein imaging, DePass detected a rare population of cDC1 dendritic cells, a cell type of intense interest in cancer immunology because of its role in priming antitumour immune responses, even though it comprised less than 0.15 percent of all cells. In a mouse embryo Spatial-mux-seq dataset spanning four modalities, the framework wove together transcriptomic, epigenomic and protein-level information into a coherent developmental picture. And on single-cell tri-modality datasets such as TEA-seq and DOGMA-seq, which measure transcripts, epitopes and chromatin accessibility simultaneously in individual cells, DePass preserved biologically meaningful dynamics, including cell-cycle progression reconstructed through pseudotime analysis.

Practical considerations may prove as important as raw performance. The authors report that DePass is scalable, employing efficient graph construction and training strategies that keep computational demands manageable even for large datasets, and they provide running time evaluations across their benchmark collection. The software is open source, with code deposited on GitHub and Zenodo, full documentation, and a detailed protocol published through protocols.io. Processed data from the study are publicly available through Zenodo and the China National Center for Bioinformation, and the paper’s data availability section catalogues accession numbers for every dataset analysed. This transparency lowers the barrier for other laboratories to adopt, test and extend the framework, which is often what determines whether a new method becomes a community standard or a footnote.

The arrival of DePass reflects a broader maturation of computational biology. A decade ago, the bottleneck in genomics was generating data; today, it is making sense of data whose complexity has outpaced the tools designed to analyse them. Multi-omics experiments hold the promise of capturing biology in its full dimensionality, but only if the layers can be fused without injecting artifacts. By building denoising into the integration process itself and by refusing to specialize in a single data type, DePass offers a template for what next-generation analysis frameworks may look like: unified, generalizable and honest about the noise inherent in real measurements. If the benchmarks hold up under independent scrutiny, the framework could become a standard workhorse in laboratories charting the molecular geography of tissues, from developing embryos to tumours on the operating table.

Subject of Research: A graph-based deep learning framework for integrating paired single-cell and spatial multi-omics data

Article Title: The dual-enhanced graph learning framework DePass allows paired data integration in single-cell and spatial multiomics

Article References: Li, W., Jiang, Y., Zhao, Q., Xu, Y., Dai, D., Rong, Y., Zhao, X., & Zhang, H. (2026). The dual-enhanced graph learning framework DePass allows paired data integration in single-cell and spatial multiomics. Nature Cell Biology. https://doi.org/10.1038/s41556-026-02067-8

Image Credits: AI Generated

DOI: 10.1038/s41556-026-02067-8

Keywords: DePass, multi-omics integration, single-cell sequencing, spatial transcriptomics, graph neural networks, deep learning, denoising, colorectal cancer, tumour microenvironment, spatial proteomics, computational biology, Nature Cell Biology

The problem the team set out to solve is a familiar one for anyone deploying artificial intelligence in the real world. Modern deep networks achieve remarkable quality by growing ever larger, but a model that runs comfortably on a cloud server is hopeless on a pair of wireless earbuds. Traditionally, engineers have trained separate models for each computational tier, or resorted to pruning and knowledge distillation, both of which require additional fine-tuning after the fact. SEFlow takes a different route: a single network is trained once in such a way that it can be dynamically sliced into subnetworks of wildly different sizes and deployed directly, with no retraining, onto anything from a smartphone to a data center.

The technical heart of the approach is a set of flexible modules the authors call FlexAttention, FlexLinear, and FlexRMSNorm. In a conventional transformer, the width of each layer is fixed: a linear layer has a set number of inputs and outputs, and multi-head attention has a fixed number of heads. FlexLinear instead slices its weight matrix to any chosen output size and, crucially, activates only a chosen subset of its input neurons, following the same logic as early-exit methods that activate only early layers. FlexAttention adjusts the number of active attention heads, and FlexRMSNorm adapts the size of its normalization parameters. Because each output is a weighted sum of many neurons, slicing works mathematically without changing what the layer computes—only how much of it runs.

Depth is handled by early exit. The network is built from a stack of residual blocks whose outputs all share the same shape, so a decoder can read the result after any number of blocks. Because of how residual connections accumulate, decoding after block number B-bar is equivalent to decoding a fusion of the features produced by the first B-bar blocks; early layers capture the main content of the signal while later layers refine details, so cutting depth sacrifices fine polish but preserves the essentials. Adjusting depth changes computation roughly linearly, while adjusting width changes it quadratically, giving the deployer a two-dimensional dial for matching any computational budget.

Training such a shape-shifting network requires care. At every training step, the team duplicates each batch: one copy trains the full network, the other trains a randomly sampled subnetwork with a random depth and width, ensuring every neuron is updated regularly. To keep multi-GPU training efficient—since GPUs assigned smaller subnetworks would otherwise finish early and sit idle—a synchronized pseudo-random generator gives every GPU the same subnetwork index each step. According to the paper, this synchronization alone cut average training computation by about 38 percent and training time by about 35 percent.

SEFlow is not merely flexible; it is also unified. Instead of training separate models for denoising, dereverberation, declipping, and packet loss concealment, the team applied dynamic data augmentation: clean speech was corrupted with random combinations of noise at signal-to-noise ratios between minus 5 and 20 decibets, simulated room reverberation, waveform clipping, and Markov-chain-simulated packet loss. One model learned to handle all of these degradations, alone or together. A lightweight auxiliary voice activity detection decoder, and a loss combining complex-spectrogram and magnitude reconstruction with the VAD objective, further boosted quality across metrics such as PESQ, STOI, and downstream speaker and speech recognition scores.

The backbone builds on BS-RoFormer, the winning system of the NeurIPS 2024 speech enhancement challenge, splitting the frequency axis into 41 sub-bands distributed approximately uniformly on the Mel scale. A two-stage band-splitting scheme makes the model sampling-rate agnostic: the same architecture handles audio at 8, 16, 22.05, 24, 32, 44.1, and 48 kilohertz simply by using fewer sub-bands at lower rates. A Distributed Grouped Sampler keeps each training batch at a single sampling rate, avoiding wasteful upsampling and GPU idle time, reducing computation by a further 17 percent in their experiments.

The results are striking in their scalability. The full network has about 27 million parameters and costs roughly 24.7 GMACs per second on 16-kilohertz audio; the smallest subnetwork, with a single residual block and a single attention head, has 1.69 million parameters and costs about 0.19 GMACs per second—two orders of magnitude less—yet still measurably improves speech quality over the unprocessed input. At full size, SEFlow performs comparably to state-of-the-art task-specific models on denoising benchmarks from the INTERSPEECH 2020 DNS Challenge and on packet loss concealment benchmarks from the 2022 PLC Challenge, with demo-quality declipping shown on the project homepage. Interesting quirks emerged: a six-block, one-head configuration beat a one-block, four-head early-exit variant across all metrics despite using only 37 percent of its compute, suggesting width can matter more than depth in some regimes.

The authors are candid about trade-offs. Flexibly trained models slightly underperform fixed full-scale networks at the same nominal size, packet loss concealment demands at least two blocks for usable performance, and automatic speech recognition accuracy downstream remains limited, hinting that the backbone or loss may need task-specific tuning. There are also overheads in training multiple subnetworks simultaneously and in deciding which subnetwork to invoke at inference. Still, the researchers argue the approach carries real environmental promise: fixed models burn peak energy even on easy inputs, whereas adaptive slicing reduces multiply-accumulate operations and memory accesses—the dominant energy costs on edge chips—potentially trimming the carbon footprint of always-on speech processing.

What makes SEFlow resonate beyond acoustics is its implication for how AI might be delivered everywhere at once. Rather than a patchwork of bespoke models scattered across earbuds, phones, cars, and servers, a single family-model system could flow intelligence across the device-edge-cloud continuum, expanding and contracting to fit each platform. If the same recipe—flexible width, early exit, unified multi-task training—transfers to language, vision, and audio generation models, the paper’s vision of ubiquitous, resource-aware intelligence moves a step closer to reality. Demonstrations of SEFlow’s outputs, from heavily noisy cafe chatter to clipped and packet-mangled calls, are publicly available, and the code is available on request, inviting the community to stress-test this elegantly elastic architecture.

Subject of Research: A flexibly scalable, unified neural architecture for multi-task speech enhancement

Article Title: Towards a flexible and unified architecture for speech enhancement

Article References: Feng, L., Zhang, C., & Zhang, X.-L. (2025). Towards a flexible and unified architecture for speech enhancement. Vicinagearth, 2(1), Article 14. https://doi.org/10.1007/s44336-025-00022-z

Image Credits: AI Generated

DOI: 10.1007/s44336-025-00022-z

Keywords: speech enhancement, SEFlow, flexible neural networks, FlexAttention, early exit, AI Flow, edge computing, slimmable networks, packet loss concealment, denoising, BS-RoFormer, resource-constrained inference