The new system, described in Frontiers in Signal Processing, achieves a classification accuracy of 99.83 percent on a balanced dataset of 14,390 chest CT images, correctly distinguishing fibrotic lungs from healthy ones with a sensitivity of 99.93 percent and a specificity of 99.72 percent. But the numbers alone tell only half the story. What sets this architecture apart is its explainability: using a visualization technique called Grad-CAM++, the model generates heatmaps that highlight exactly which regions of the lung drove its decision, giving radiologists a window into the machine’s reasoning rather than forcing them to trust an opaque verdict.

The engineering challenge the team faced was fundamental to how fibrosis manifests. Unlike tumors or fractures, pulmonary fibrosis announces itself through slight, diffuse, and seemingly random texture changes scattered across the lung parenchyma. Conventional convolutional neural networks like ResNet50V2, the backbone of the new system, rely on fixed receptive fields and hierarchical feature extraction — a design that struggles to capture both the fine-grained fibrotic textures and the long-range contextual patterns the disease produces. Worse, without explicit spatial attention, such models can be distracted by irrelevant structures like ribs and background noise, contributing to the false positives that have plagued earlier automated approaches.

The researchers’ solution is a module they call Pyramid Spatial Atrous Channel Attention, or PSACA, grafted onto the ResNet50V2 transfer learning framework. The design reorders the logic of attention itself: spatial attention is applied first, directing the network’s focus to disease-relevant lung regions before anything else happens. Then comes Atrous Spatial Pyramid Pooling, a clever technique built on parallel dilated convolutions with dilation rates of 1, 5, 10, and 15. Lower rates capture intricate structural details such as edges and vascular patterns, while higher rates encode broader contextual information about patch-level and global lung architecture. Finally, channel attention selectively amplifies the most informative feature channels, completing a three-stage refinement pipeline that transforms raw CT data into a rich, fibrosis-aware representation.

Placement matters as much as design. The PSACA block was embedded at Stage 3 and Stage 4 of ResNet50V2 — the levels where the network learns mid-level structural features like localized fibrosis textures and high-level semantic patterns of lung abnormalities. Earlier stages, which capture only edges and simple textures, were left untouched, keeping computational complexity low while maximizing discriminative power where it counts. An ablation study confirmed the wisdom of this strategy: the baseline ResNet50V2 achieved 98.61 percent accuracy, but adding PSACA at Stage 3 or Stage 4 individually pushed performance to 99.69 percent, and combining both stages yielded the full 99.83 percent. The components, the authors found, contribute complementary rather than redundant information.

The team began by benchmarking eight pre-trained deep learning architectures — VGG19, MobileNetV2, ResNet-50, ResNet50V2, ResNet-152V2, DenseNet-121, DenseNet-169, and DenseNet-201 — all initialized with ImageNet weights and fine-tuned on the fibrosis dataset. ResNet50V2 emerged as the strongest performer, and the researchers then customized it further, unfreezing the final 20 layers, adding a global average pooling layer, a dense layer of 512 neurons with ReLU activation, and a 50 percent dropout layer to guard against overfitting before a final sigmoid output for binary classification. Training used the Adam optimizer with a learning rate of 0.001, binary cross-entropy loss, a batch size of 32, and 30 epochs — hyperparameters identified through Bayesian optimization, which the authors found converged faster than grid or random search.

The dataset itself, sourced from Kaggle, was carefully balanced: 7,195 normal CT images and 7,195 fibrosis images, split 60/20/20 into training, validation, and test sets. On the held-out test data, the model missed just one fibrotic scan out of 1,439 and misclassified only four normal scans. The receiver operating characteristic curve told an even more striking story, with an area under the curve of 1.00 — a perfect equilibrium between sensitivity and specificity that, if it holds up under independent scrutiny, would represent a remarkable discriminatory ability for any diagnostic tool.

The explainability layer may prove to be the system’s most clinically valuable feature. Grad-CAM++ heatmaps, superimposed on lung-only CT images after segmentation, consistently lit up fibrosis-affected tissue along lung borders, with green marking affected regions and red flagging highly confident lesion cores. This weakly supervised localization allows radiologists to verify that the model is attending to genuine pathological features rather than artifacts — a critical safeguard for clinical trust. The authors note that such visual validation enhances radiologist confidence and addresses the transparency deficits that have hindered AI adoption in interstitial lung disease imaging, a concern echoed across recent reviews of machine learning in radiology.

Compared with prior approaches — from a multiscale rotation-invariant CNN that reached 90 percent accuracy to a CNN-transformer parallel fusion model at 97.6 percent — the PSACA-ResNet50V2 system represents a substantial leap. Yet the authors are candid about its limits. The model was trained and evaluated on a single publicly available dataset, and real-world clinical data varies widely in acquisition protocols, scanner types, and patient demographics. Validation on multi-institutional datasets is the necessary next step before the system can move toward real-time diagnostic deployment, and the researchers suggest extending the framework to multi-class classification of multiple lung diseases.

Still, the study offers a compelling glimpse of where medical AI is heading: not toward ever-larger black boxes, but toward architectures engineered from the ground up to see the way clinicians see — attending to subtle textures, weighing context across scales, and explaining themselves in images a radiologist can read. For a disease where median survival is measured in years and diagnosis is measured in missed opportunities, an algorithm that can both detect fibrosis with 99.83 percent accuracy and show exactly where it looked may be exactly the kind of transparent ally pulmonary medicine has been waiting for.

Subject of Research: Explainable deep learning model for pulmonary fibrosis diagnosis from chest CT images

Article Title: Multi-scale pyramid spatial atrous channel attention enhanced ResNet50V2 for explainable pulmonary fibrosis diagnosis

Article References: Mahapackialakshmi, K., & Jaffino, G. (2026). Multi-scale pyramid spatial atrous channel attention enhanced ResNet50V2 for explainable pulmonary fibrosis diagnosis. Frontiers in Signal Processing, 6, Article 1787092. https://doi.org/10.3389/frsip.2026.1787092

Image Credits: AI Generated

DOI: 10.3389/frsip.2026.1787092

Keywords: pulmonary fibrosis, deep learning, medical imaging, explainable AI, Grad-CAM++, ResNet50V2, attention mechanisms, chest CT, artificial intelligence, diagnostic AI, Multi-scale, pyramid

The core insight behind the work is that clinical photographs and dermoscopic images, while complementary, carry fundamentally different kinds of information. Dermoscopic images expose detailed morphological features such as pigment networks, globules, and vascular patterns that are central to differential diagnosis. Clinical images, by contrast, situate the lesion within its broader anatomical and visual context but offer less diagnostic resolution at the cellular and structural level. Most existing multimodal fusion approaches in medical imaging treat both modalities symmetrically, blending their features with equal weight as if neither view were inherently more diagnostic. Chu and Yang argue that this symmetry is a mistake, and they propose an alternative they call Dermoscopy-Dominant Correction Fusion, or DDCF.

DDCF is an asymmetric fusion framework. Rather than merging clinical and dermoscopic features as equals, it preserves the dermoscopic representation as the base, the primary backbone of the model’s understanding of the lesion. Clinical information then enters not as an equal partner but as a gated corrective signal, a mechanism that modulates and refines the dermoscopic features without overwhelming them. This design choice reflects the clinical reality that dermoscopy is the more diagnostically powerful modality for skin lesion classification, and that clinical context should serve to correct or contextualize rather than to dominate. The framework was tested across three different neural network backbones: Swin-Tiny, a vision transformer variant; DenseNet121, a densely connected convolutional network; and EfficientNet-B4, a compound-scaled convolutional architecture. Testing across multiple backbones is important because it helps distinguish improvements attributable to the fusion strategy itself from improvements attributable to any single model’s particular inductive biases.

The evaluation was conducted on the MILK10k dataset, using nine diagnostic categories that span a clinically meaningful range of skin lesions. The researchers created three stratified lesion-level partitions with a 7:1:2 training, validation, and test ratio, holding class-wise counts fixed across splits. This stratification ensures that rare diagnostic categories are not underrepresented in any partition, a critical concern in medical imaging where class imbalance can silently inflate or deflate performance metrics. For each backbone and each partition, the correction coefficient was fixed at a specific value, and the model checkpoint with the lowest validation loss was selected before any test evaluation was performed. This protocol guards against overfitting to the test set and provides a more honest estimate of real-world generalization.

The primary evaluation metric was macro-averaged one-vs-rest area under the receiver operating characteristic curve, or AUROC, which measures the model’s ability to rank positive cases above negative ones across all diagnostic categories regardless of their prevalence. Balanced accuracy and macro-F1 score served as complementary metrics. To compare DDCF against a baseline, the researchers used naive-sum fusion, a straightforward approach that simply adds the feature representations from both modalities without any gating or asymmetry. Statistical comparison was performed using paired stratified bootstrap confidence intervals within each backbone and partition combination, yielding nine total backbone-by-partition comparisons.

The results paint a nuanced picture. At a correction coefficient of 0.7, DDCF achieved mean test AUROCs of 0.9308 for Swin-Tiny, 0.9099 for DenseNet121, and 0.9084 for EfficientNet-B4. This coefficient produced the highest observed mean test AUROC among all evaluated values for all three backbones, suggesting that a moderate but clearly dermoscopy-dominant weighting is optimal. In all nine backbone-by-partition comparisons, the AUROC difference between DDCF and naive-sum fusion was positive, meaning that the asymmetric approach numerically outperformed the symmetric baseline every single time. In two of those nine comparisons, the nominal 95 percent bootstrap confidence intervals excluded zero, providing formal statistical support for the improvement in those specific configurations. Macro-F1 differences were also positive across all nine comparisons, indicating that the gains were not limited to a single metric.

However, the study is careful not to overstate its findings. Balanced accuracy told a more complicated story: for Swin-Tiny, balanced accuracy was actually lower with DDCF than with naive-sum fusion in all three partitions, with a mean of 0.5736 versus 0.6229 for the baseline. This discrepancy between AUROC and balanced accuracy highlights an important tension in medical AI evaluation. AUROC measures ranking ability across thresholds, while balanced accuracy depends on a specific operating point, and a model can improve in one while declining in the other. The researchers also found that class-wise improvements were particularly evident for benign keratosis-like lesions, abbreviated BKL, and basal cell carcinoma, abbreviated BCC. Yet even here, the increased sensitivity for BCC came at the cost of reduced specificity, meaning the model became better at catching BCC cases but also more prone to false positives for that category.

Perhaps most notably, the authors found no uniform benefit for melanoma or for keratinocytic tumors as a whole. This is a significant caveat given that melanoma detection is often the primary motivation for advancing skin lesion classification systems. The results do not establish a universally optimal correction coefficient, nor do they demonstrate general clinical superiority of DDCF over simpler fusion approaches. What they do establish is a consistent directional signal: the asymmetric, dermoscopy-dominant approach produced positive numerical AUROC differences in every tested configuration, which is a meaningful pattern even if individual comparisons did not all reach conventional statistical significance.

The work is funded by the National Research Foundation of Korea through the Ministry of Science and ICT and represents a growing recognition that how multimodal information is combined in medical AI is as important as which modalities are used. Rather than treating fusion as a solved problem, Chu and Yang demonstrate that the architecture of combination itself, specifically which modality serves as the foundation and which serves as a corrective influence, can measurably affect diagnostic performance. The study analyzed a publicly available, de-identified dataset and did not enroll new human participants, and the authors declare no competing interests. As multimodal medical AI continues to mature, findings like these suggest that the field may benefit from moving beyond symmetric fusion paradigms and toward designs that explicitly encode the clinical hierarchy of diagnostic evidence, letting the most informative modality lead while the other provides context and correction.

Subject of Research: Asymmetric multimodal fusion of clinical and dermoscopic images for AI-based skin lesion classification

Article Title: Dermoscopy-dominant correction fusion for paired clinical and dermoscopic skin lesion image classification

Article References: Chu, Y., & Yang, S. (2026). Dermoscopy-dominant correction fusion for paired clinical and dermoscopic skin lesion image classification. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02955-3

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02955-3

Keywords: dermoscopy, skin lesion classification, multimodal fusion, medical imaging, artificial intelligence, deep learning, clinical images, AUROC, melanoma, basal cell carcinoma, computer vision, diagnostic AI

The study, published in Nature Machine Intelligence by Luyang Cao, Zhijian Wu and colleagues at Nanjing University, Westlake University, Southeast University and the Shenzhen Institute of Advanced Technology, was trained and evaluated on an extraordinary corpus: 3.76 million paired images, in which each degraded clinical image is matched with a high-quality reference. The data were assembled from publicly available repositories including The Cancer Imaging Archive, OpenNeuro, NITRC, Grand Challenge, Synapse, Kaggle and GitHub, and span 34 MRI datasets covering brain, cardiac, abdominal and prostate imaging, 37 CT datasets ranging from thoracic to whole-body acquisitions, six X-ray datasets, five endoscopy datasets, four fundus photography datasets, two ultrasound datasets and one whole-body PET dataset for oncological imaging.

Architecturally, AMIFound follows a philosophy that has transformed natural-language processing but has been slower to arrive in medical image restoration: a shared backbone with adaptive, task-specific routing. A single encoder first maps any low-quality input, regardless of modality, into a common latent feature space that captures both modality cues and the nature of the degradation. A dynamic routing mechanism then selects from a pool of specialised expert modules to construct a restoration pathway tailored to each individual image. The expert pool contains fast Fourier transform attention experts that operate in the frequency domain, spatial attention experts, convolutional local-feature experts and hybrid FFT-CNN fusion experts, allowing the model to combine complementary frequency, spatial and structural information on the fly.

Training such a system required a way to generate realistic degradation across modalities. The researchers built a modality-aware degradation engine that transforms high-quality images into degraded observations using physics-informed simulations. For CT, metal artefacts are synthesised in the projection domain through random metal masks, Radon transformation, nonlinear artefact mapping and inverse reconstruction. MRI aliasing and low-resolution effects are produced by k-space truncation followed by inverse Fourier reconstruction, while PET low-dose degradation is simulated with Poisson noise, Gaussian noise and blur. Endoscopic low-light conditions are modelled by gamma correction and brightness reduction in HSV colour space, X-ray blur by cascaded filtering with anisotropic diffusion, ultrasound artefacts by reverberation echoes, geometric shadowing and depth attenuation, and fundus artefacts by random circular highlights and Gaussian softening. In total, the model handles seven representative degradation patterns.

The benchmark results are striking. Across all seven modalities, AMIFound outperformed 38 competing baselines, including both task-specific restoration networks and generalist models such as AMIR, MoCE-IR and Found-IR, on full-reference image quality metrics such as structural similarity and learned perceptual measures. Radar plots of structural similarity improvements show consistent gains over each competing framework, and detailed per-task comparisons show the model achieving the best or near-best scores against modality-specific methods for every restoration task, from CT metal artefact reduction to PET denoising and cataract-affected fundus enhancement.

Quantitative metrics alone do not convince radiologists, so the team also ran a subjective evaluation with clinicians across all seven modalities. Physicians rated the restored images on sharpness, artefact suppression, naturalness and diagnostic confidence. AMIFound achieved the highest or near-highest average scores in every modality and provided greater diagnostic confidence than competing methods. The model was also stress-tested under combined degradations, where multiple corruptions strike the same image simultaneously, and on naturally degraded clinical images rather than synthetic ones, maintaining its advantage in both settings and preserving anatomical structures and realistic intensity distributions where competitors tended to oversmooth or distort contrast.

Perhaps the most clinically consequential finding is what happens downstream. The researchers fed AMIFound-enhanced images into 44 disease detection tasks and found consistent improvements over detection on the original degraded images. The benefits extended to tumour stage prediction in non-small-cell lung cancer and to survival risk stratification, including improved individual-level prognostic ranking. Attention heatmaps from the CT-based tumour-stage classifier reveal a plausible mechanism: before enhancement, model attention is often diffuse or mislocalized and predictions are less certain, whereas after enhancement, attention concentrates on tumours, lung-parenchyma abnormalities and lesion boundaries, and the highest prediction probability more often agrees with the ground-truth stage.

The significance of this work lies partly in its economics. Hospitals generate enormous volumes of routine imaging that falls short of research-grade quality, and low-dose protocols, deliberately chosen to reduce radiation exposure, or motion artefacts from patients who cannot hold still, are unavoidable facts of clinical life. Previous approaches required a separate, carefully tuned network for each modality-degradation combination, an unsustainable proliferation as imaging technology diversifies. A single foundation model that generalises across modalities and degradations, trained once and deployed everywhere, could lower the barrier for smaller institutions and make image enhancement a routine preprocessing step for any AI diagnostic pipeline.

The authors have released their training and evaluation code publicly on GitHub, with a versioned archive on Zenodo, built on PyTorch and drawing on data-loader pipelines from prior all-in-one restoration frameworks such as PromptIR, AirNet and MoCE-IR. All datasets used were publicly available and accessed under their respective licences, and the team reports that the model preserves clinically relevant information rather than hallucinating plausible-looking detail, a critical concern for any generative approach in medicine. Funded by the National Natural Science Foundation of China and several regional programmes, the work positions AMIFound as a domain-specific foundation model for the unglamorous but essential task of making everyday clinical images good enough, not just for human eyes, but for the growing fleet of AI systems that depend on them.

Subject of Research: A multimodal foundation model for restoring and enhancing degraded medical images across seven imaging modalities

Article Title: An all-in-one foundation model for multimodal medical image restoration and enhancement

Article References: Cao, L., Wu, Z., Zhang, J., Qi, L., Gao, Y., Zheng, H., Zheng, Y., & Shi, Y. (2026). An all-in-one foundation model for multimodal medical image restoration and enhancement. Nature Machine Intelligence. https://doi.org/10.1038/s42256-026-01310-z

Image Credits: AI Generated

DOI: 10.1038/s42256-026-01310-z

Keywords: medical imaging, foundation model, image restoration, deep learning, AMIFound, CT, MRI, PET, ultrasound, endoscopy, lung cancer, diagnostic AI

To understand why the review matters, it helps to grasp what makes an AI system agentic in the first place. Most pathology AI deployed today is a fixed pipeline: an image goes in, a prediction comes out, and the model has no ability to change its own behavior at inference time. Agentic systems break that mold. They coordinate multiple components, including perception models that analyze images, large language models that reason over textual knowledge, external tools such as databases or specialized classifiers, and feedback loops that let the system revise its own intermediate conclusions. Crucially, the control flow is dynamic rather than predetermined. The system decides, moment by moment, which tool to call, which region of a slide to examine next, and when it has gathered enough evidence to render a judgment.

The authors organize this emerging field with an operational taxonomy built on three pillars: dynamic control flow, inference-time tool selection, and knowledge integration. This framework matters because the term agentic has become a marketing buzzword as much as a technical one, applied to everything from simple chatbot wrappers to genuinely autonomous multi-model pipelines. By defining agency in terms of what a system actually does at inference time, whether it selects its own tools, integrates new knowledge on the fly, and adapts its reasoning path, the review gives researchers and regulators a common vocabulary for evaluating claims. Under this taxonomy, the authors survey architectures and enabling technologies across the major application domains of computational pathology: diagnosis, prognosis, and therapeutic support.

The technical appeal of agentic designs is easy to see. A single static model trained to detect one cancer type in one organ struggles the moment it encounters an unusual case, a rare variant, or a question it was never trained to answer. An agentic system, in principle, could recognize the limits of its own competence, consult a reference knowledge base, dispatch a specialized sub-model for a second opinion, and synthesize the results into a coherent report, much as a pathologist might consult a colleague or the literature. Early prototypes demonstrate that this kind of flexible orchestration is feasible, and the review treats that feasibility as established. The perception components, the language models, and the orchestration layers all exist and can be wired together into functioning systems.

Feasibility, however, is where the good news ends, at least for now. The review’s central and most consequential finding is that agentic architectures have not yet demonstrated clinical benefit. The evidence base consists almost entirely of retrospective benchmarks and research prototypes, studies in which systems are tested on archived data rather than deployed in live diagnostic workflows. Retrospective performance, the authors emphasize, is a notoriously poor predictor of real-world clinical value. A model that achieves impressive accuracy on a curated test set may fail in unpredictable ways when confronted with the staining variability, scanning artifacts, and case mix of an actual hospital.

There is a deeper methodological problem lurking in the literature, and the review names it directly: reported gains are difficult to attribute to agentic organization itself. When a new agentic system outperforms an older one, is the improvement due to the clever orchestration, or simply to a stronger underlying vision model, more training data, or a larger inference budget that lets the system think longer? Because studies differ in backbone models, datasets, and computational resources, most published comparisons cannot isolate the contribution of agency. This is a familiar trap in AI research, where architectural novelty often gets credit for gains that actually come from scale. Until head-to-head comparisons are run with matched backbones, matched data, and matched compute, the field cannot honestly claim that agentic organization, as opposed to brute capability, is what delivers the benefit.

The review also catalogues the specific risks that agentic designs introduce on top of those inherent in any medical AI. Large language models can hallucinate, generating confident but false statements, and in a pathology context a hallucinated finding woven into a diagnostic report is a patient-safety hazard. Multi-component systems expand the attack surface for security threats, since each tool call and knowledge lookup is a potential point of failure or manipulation. Computational cost is another practical concern: agentic systems that invoke multiple models and iterative reasoning loops consume far more compute than a single forward pass through a classifier, which affects both latency and the economics of deployment. And workflow integration remains the perennial stumbling block for pathology AI, because a system that is accurate but awkward to use, or that disrupts how pathologists actually work, will simply go unused.

Regulatory and human factors add further layers of complexity. Dynamic systems whose behavior depends on inference-time decisions are harder to validate and lock down than fixed models, raising unresolved questions about how regulators should assess software that does not always do the same thing given the same input. The authors also point to patient preferences as a consideration that is often absent from technical papers but central to clinical adoption, alongside the non-negotiable requirement that pathologists retain meaningful oversight of any AI-assisted diagnosis. Notably, the review does not treat agentic AI as inevitable. It explicitly identifies conditions under which specialist non-agentic models remain preferable, a stance that cuts against the industry narrative that autonomy is always the goal. For a well-defined task such as detecting mitotic figures or grading a common tumor subtype, a narrow, validated, inexpensive model may simply be the better tool.

So what would it take for agentic pathology AI to earn clinical trust? The authors lay out a translation roadmap that reads as a checklist for the field. First, prioritize verifiable tasks, meaning applications where the correctness of the AI’s output can be objectively checked, rather than open-ended reasoning with no ground truth. Second, run matched comparisons that isolate the effect of agentic organization from confounds like model scale. Third, pursue prospective and external validation, testing systems on new cases, at new institutions, in real time, rather than re-splitting old datasets. Fourth, establish lifecycle governance, continuous monitoring and updating frameworks appropriate to systems that may behave differently over time as their components evolve. Finally, design interfaces that preserve pathologist oversight, keeping the human diagnostician in command rather than relegating them to rubber-stamping machine output.

The timing of this sober assessment is significant. Agentic AI is arguably the hottest concept in the technology sector, and medicine is under intense pressure to adopt it, with vendors promising autonomous diagnostic agents that could ease workforce shortages and accelerate care. This review, published open access and grounded in a systematic survey rather than promotional claims, offers a counterweight: a technically literate account that concedes the engineering achievements while insisting that the clinical evidence simply does not exist yet. For pathologists, hospital administrators, and regulators, the message is to distinguish enthusiasm from evidence. For researchers, it is a call to design the rigorous studies that could eventually justify deployment. Agentic systems may well reshape computational pathology, but as this analysis makes clear, the road from impressive demo to dependable diagnosis runs through prospective trials, honest comparisons, and governance structures that keep patients, and the pathologists who serve them, firmly in the loop.

Subject of Research: Agentic AI architectures for computational pathology and their translational challenges

Article Title: Agentic systems in computational pathology: architectures, evidence, and translational challenges

Article References: Lu, X., Li, Q., Gao, Y., Dong, W., Lyu, M., Ma, S., Li, J., Wu, Y., Cai, L., Zhang, T., Lyu, S., Liu, Z., Liu, H., & Luo, S. (2026). Agentic systems in computational pathology: architectures, evidence, and translational challenges. Journal of Translational Medicine, 24(1), Article 1111. https://doi.org/10.1186/s12967-026-08829-0

Image Credits: AI Generated

DOI: 10.1186/s12967-026-08829-0

Keywords: agentic AI, computational pathology, digital pathology, whole-slide imaging, large language models, clinical validation, translational medicine, AI regulation, hallucination risk, diagnostic AI, prospective validation, pathologist oversight

The study, published in PLOS Neglected Tropical Diseases, is a scoping review led by Rose Mary Maniraj and colleagues at Universiti Kebangsaan Malaysia, working with researchers at Malaysia’s National Institutes of Health. The team systematically searched seven electronic databases for studies published between January 2015 and July 2025, a decade chosen because it marks the era when convolutional neural networks began to dominate biomedical image analysis. After screening more than 1,000 records and supplementing the search with backward citation tracking, the researchers assembled 118 primary studies that applied deep learning models to parasite microscopy images for human clinical diagnosis and reported quantitative performance metrics.

The verdict on raw technical capability is striking. Across the included studies, deep learning models consistently achieved high performance in detecting and classifying parasites, with some systems matching the accuracy of trained microscopists under controlled conditions. Yet the review’s authors caution that these headline numbers conceal a fragmented and uneven evidence base. Studies varied enormously in datasets, architectures, evaluation metrics and experimental designs, making direct comparison of one model against another essentially impossible. More troubling, most studies validated their models internally — on subsets of the very same data used for training — rather than testing them on genuinely independent patient populations, staining protocols or laboratory conditions.

The technical landscape mapped by the review is dominated by three computer vision tasks. Image classification, in which a model decides whether an image contains a parasite and often which species, was the most common, appearing in 69 of the 118 studies. Object detection, which goes further by drawing bounding boxes around individual parasites and enabling quantification of infection burden, accounted for 43 studies. Image segmentation — pixel-level delineation of parasite structures that permits detailed morphological analysis — was the rarest, with just six studies. The scarcity of segmentation work likely reflects the enormous labor required to produce pixel-accurate annotations, a bottleneck familiar to anyone who has worked in medical image labeling.

Malaria looms large over the field. Plasmodium species, the parasites responsible for a disease that caused an estimated 263 million cases and nearly 600,000 deaths in 2023, dominated both the classification and detection literature, driven in part by the availability of large, publicly labeled datasets of stained blood smear images. Classification studies leaned heavily on convolutional neural network architectures — VGG, ResNet, DenseNet, EfficientNet and custom variants — often enhanced with attention mechanisms, transfer learning and, increasingly, transformer-based designs. Detection studies overwhelmingly favored the YOLO family of single-shot detectors, spanning versions from YOLOv2 through YOLOv11, alongside lightweight adaptations engineered for speed and low computational cost.

Beyond malaria, the review catalogued deep learning applications across a striking breadth of parasites. Protozoan targets included Babesia, Giardia lamblia, Trypanosoma cruzi, Toxoplasma gondii, Trichomonas and Leishmania species, the latter detected in bone marrow and cutaneous lesion microscopy. Helminth studies covered Schistosoma haematobium and S. mansoni eggs, Ascaris lumbricoides, Taenia saginata, Opisthorchis viverrini, Trichuris trichiura, microfilariae and the broader category of soil-transmitted helminths. Detection frameworks such as SSD MobileNet, Faster R-CNN, RetinaNet and even transformer-based detectors like RT-DETR appeared alongside the YOLO mainstream, while segmentation efforts centered on U-Net and its many variants.

What sets this review apart from earlier syntheses of artificial intelligence in parasitology is its explicit focus on implementation — the unglamorous question of whether any of these algorithms can actually function in a working diagnostic laboratory. The answer, mostly, is not yet. Several studies proposed integration pathways: deep learning models serving as preliminary screening tools that flag suspicious samples for human review, outputs feeding into clinical decision support systems, or deployment on smartphone-based microscopes and single-board computers such as the Raspberry Pi for point-of-care use. Others sketched visions of cloud-based platforms, Internet of Medical Things architectures and telemedicine frameworks that would allow remote expert consultation in resource-limited settings.

A handful of studies did venture beyond proof-of-concept. The review highlights real-time and near real-time implementations, including smartphone applications for automated malaria screening, mobile digital microscopy paired with deep learning for detecting soil-transmitted helminths and schistosome eggs in field settings, and a semi-automated workflow combining whole-slide scanning, AI analysis and expert verification that achieved rapid turnaround times. One study compared AI-supported microscopy against manual reading of Kato-Katz smears in a primary healthcare setting. But these remain exceptions. The overwhelming majority of the 118 studies never left the experimental bench, and considerations such as usability, scalability, cost-effectiveness and regulatory compliance were rarely addressed.

The authors are candid about the limitations of their own analysis. Only English-language articles were included, no formal critical appraisal of study quality was performed — consistent with scoping review methodology — and the reliance on published performance metrics raises the specter of publication bias, since studies reporting favorable results are more likely to see print. They also note a representation problem: the heavy reliance on publicly available malaria datasets may skew the field toward well-studied parasites, leaving neglected infections underrepresented. Dataset sizes, moreover, were reported in incompatible units — whole-slide images, fields of view, individual cells, image patches — further muddying any attempt at quantitative synthesis. No formal meta-analysis was attempted, and the review’s performance findings are presented descriptively rather than pooled.

The bottom line is a field brimming with promise but facing a translational gap. Deep learning has demonstrated that it can see parasites — reliably, quickly and, in some architectures, cheaply enough to run on a phone. What it has not yet demonstrated, in most cases, is that it can survive contact with the messy realities of clinical parasitology: variable staining, ambiguous morphology, unfamiliar patient populations and overworked laboratory staff. The review’s authors call for standardised datasets, rigorous external validation and prospective, implementation-focused studies that measure not just accuracy but operational feasibility and clinical utility. Until then, the gold standard remains a trained human eye — increasingly assisted, but not yet replaced, by the machine.

Subject of Research: Deep learning applications for automated parasite detection and diagnosis in clinical microscopy

Article Title: Deep learning applications in parasite microscopy: A scoping review

Article References: Maniraj, R. M., Osman, E., Sheikh Abdullah, S. N. H., Azil, A. H., & Mahmud, M. A. F. (2026). Deep learning applications in parasite microscopy: A scoping review. PLOS Neglected Tropical Diseases, 20(10), e0014774. https://doi.org/10.1371/journal.pntd.0014774

Image Credits: AI Generated

DOI: 10.1371/journal.pntd.0014774

Keywords: deep learning, parasitology, microscopy, malaria, convolutional neural networks, object detection, image classification, image segmentation, diagnostic AI, neglected tropical diseases, scoping review, clinical validation

The study stands out for a methodological choice that most clinical AI deployments still lack. Conventional binary classifiers — the workhorses of medical machine learning — produce a single probability for every case, and a fixed threshold converts that probability into a decision of malignant or benign. Such systems never signal doubt. They output a number even when the input image is ambiguous, the acquisition quality is poor, or the case falls far outside anything they were trained on. The research team, led by first authors Hang Ling, Cailing Lin and Jing Ning, with corresponding author Ziwei Zhang, built their framework around a different premise: that a diagnostic algorithm should be able to abstain. Their selective triage rule combines three ingredients — a fusion prediction model, a composite uncertainty score, and a pre-locked routing policy that sends confident cases down automatic pathways and routes uncertain ones to senior human review.

The technical architecture is worth unpacking. The fusion model integrates three distinct information streams: structured clinical data, features extracted from multimodal ultrasound imaging, and variables mined from structured ultrasound reports. Multimodal ultrasound is an important component here, since modern neck ultrasound is not a single image but a constellation of modalities — grayscale morphological features such as echogenicity and border characteristics, Doppler vascular patterns, and in some settings contrast-enhanced ultrasound dynamics. By fusing these with clinical context and report-derived information, the model aims to approximate the holistic judgment a seasoned sonographer applies, rather than relying on pixel patterns alone.

Just as important is what the researchers did with uncertainty. Rather than trusting the model’s raw confidence, they constructed a composite uncertainty measure from two complementary signals: predictive entropy, normalized against its distribution in the development cohort, and probability dispersion estimated through bootstrap resampling of the model’s fits. Entropy captures how peaked or flat the model’s output distribution is on a given case, while bootstrap dispersion captures how sensitive the prediction is to perturbations in the training data — a proxy for how far the case sits from the model’s comfort zone. Combining the two, the team fixed an uncertainty threshold of U = 0.700. Critically, every free parameter was locked during development: the binary classification threshold at p = 0.537, the selective-routing probability boundaries at p = 0.320 and p = 0.660, and the uncertainty cutoff. Once locked, the entire pipeline was applied to the internal-validation and two external-validation cohorts without any retuning whatsoever — a design that guards against the subtle overfitting that plagues many retrospective AI studies.

The study population comprised 518 patients, with one index lymph node analyzed per person: 206 in the development cohort, 88 in internal validation, and 112 in each of two independent external cohorts. Discrimination remained remarkably stable across sites, a finding that in itself deserves attention because performance degradation at external sites is the most common failure mode of published clinical prediction models. The area under the receiver operating characteristic curve was 0.839 in internal validation, 0.850 in external validation cohort 1, and 0.842 in external validation cohort 2. At the locked binary threshold, sensitivity and specificity were 0.595 and 0.902 in internal validation, 0.657 and 0.911 in the first external cohort, and 0.817 and 0.808 in the second. Those numbers describe a competent but not extraordinary classifier — which is precisely the point, because the selective framework was engineered to compensate for the model’s fallibility rather than to hide it.

The heart of the paper lies in its selective-triage results. In the pooled external-validation population of 224 patients, 38 were routed to a lower-risk automatic pathway, 44 to a higher-risk automatic pathway, and 142 — more than sixty percent — were deferred for senior review. Automatic coverage was therefore 0.366, while selective accuracy among the automated cases reached 0.927, with a 95 percent confidence interval of 0.849 to 0.966. The accepted errors within the automated subset were small in absolute terms: two false negatives and four false positives. Among patients funneled into the lower-risk automatic pathway, the negative predictive value was 0.947, meaning the residual probability of malignancy in that group was 5.3 percent — a figure the authors report transparently with a wide confidence interval spanning roughly 1.5 to 17.3 percent. The higher-risk pathway achieved a positive predictive value of 0.909. A risk–coverage analysis, summarized by a partial area under the risk–coverage curve of 0.037 in the pooled external data, quantified how selective accuracy behaved as coverage was expanded or contracted.

The authors are unusually explicit about the limits of these numbers. Only about 37 percent of external cases were eligible for automatic routing, and the accepted false-negative and false-positive counts, while modest, are not zero. Two of the 38 patients automatically assigned to the lower-risk pathway turned out to have malignant nodes. In a separate per-case analysis, the composite uncertainty score was only moderately effective at flagging binary-model errors, achieving an area under the curve of 0.581 for error detection and an area under the precision-recall curve of 0.247 — numbers that indicate real room for improvement in the uncertainty estimation itself. The team concludes that the framework should be interpreted as a retrospective, research-stage selective-routing demonstration rather than an established clinical safety or workflow tool, a disclaimer that rare in a field where press releases routinely outrun the evidence.

Where the system showed its most immediately practical benefit was in the eight-reader study. Using crossed reader-by-case bootstrap resampling — the gold-standard multi-reader multi-case methodology for interpreting diagnostic accuracy studies — the researchers tested whether access to the model’s output changed reader performance. Junior readers improved their accuracy by 0.055, a statistically significant gain with a confidence interval of 0.022 to 0.089 and a p-value of 0.002. Middle-level and senior readers showed no statistically significant change, a pattern consistent with the intuition that the tool functions as a form of expert guidance for less experienced practitioners while adding little for those who already possess the pattern-recognition skills it encodes. If the framework ultimately translates to the clinic, its clearest value proposition may be compressing the training gap between junior and senior diagnosticians, rather than replacing expert judgment outright.

The broader significance of the study lies in its modeling of what responsible clinical AI could look like. Rather than chasing headline accuracy figures, the researchers built their entire evaluation around the questions that actually matter for deployment: when should an algorithm be allowed to act autonomously, how much residual risk is acceptable within the automated zone, and how much workload is genuinely deferred to humans. The reported numbers answer those questions soberly. Roughly a third of cases could be automated with a combined accuracy above 92 percent, but the framework’s own honesty mechanisms pushed nearly two-thirds of patients toward senior review, and even the automated pathway retained a nonzero malignancy risk. The team also adhered to modern reporting and risk-of-bias standards, referencing the TRIPOD+AI and PROBAST+AI frameworks, and the study received ethics approval from Fujian Provincial Hospital with center-specific authorization for the external cohorts. Funded by the Natural Science Foundation of Fujian Province, the work offers a template — conservative, externally validated, and uncertainty-aware — for a generation of diagnostic AI systems whose most important capability may be knowing what they do not know.

Subject of Research: An uncertainty-aware selective artificial intelligence triage model for distinguishing benign from malignant cervical lymphadenopathy on multimodal ultrasound

Article Title: An uncertainty-aware ultrasound triage model for cervical lymphadenopathy: a retrospective multicenter development and external validation study

Article References: Ling, H., Lin, C., Ning, J., Liu, X., Dong, C., & Zhang, Z. (2026). An uncertainty-aware ultrasound triage model for cervical lymphadenopathy: a retrospective multicenter development and external validation study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02780-8

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02780-8

Keywords: cervical lymphadenopathy, ultrasound, artificial intelligence, uncertainty quantification, selective prediction, multimodal imaging, external validation, triage, diagnostic AI, reader study, medical imaging, clinical decision support