The research, conducted by Ahmad Ryad Soobhany and Mohammed Thafheem at Heriot-Watt University in Dubai, tackles one of the most stubborn bottlenecks in medical imaging. While it has become relatively routine for deep learning models to answer a binary question—is this scan diseased or not?—assigning a disease stage requires expert annotation that is expensive, slow, and prone to disagreement between specialists. Public datasets rarely include severity labels for optical coherence tomography, the cross-sectional imaging technique that has become the workhorse of retinal diagnosis. The team’s answer was to sidestep the problem entirely, letting an algorithm discover severity patterns on its own.

The stakes are considerable. According to the World Health Organization, AMD is a leading cause of vision loss in developed countries, and roughly 200 million people worldwide already live with some form of the disease. The global population of older adults is projected to swell from around 962 million to two billion by 2050, which means the number of OCT scans requiring expert interpretation will only grow. Because AMD has no cure, the window for meaningful intervention lies in its early stages, when deposits called drusen—yellow accumulations beneath the retina—can be monitored before they progress to vision-threatening disease.

AMD comes in two broad forms. Dry AMD, which accounts for 80 to 85 percent of cases, progresses slowly as drusen accumulate and the retinal pigment epithelium thins. Wet AMD, affecting the remaining 15 to 20 percent, is far more aggressive: abnormal blood vessels grow beneath the retina in a process called choroidal neovascularization, leaking fluid and blood into the macula and causing rapid, severe vision loss. Roughly 10 to 15 percent of dry cases eventually convert to the wet form, which is precisely why reliable early detection and staging matter so much—catching the disease before conversion gives clinicians a chance to act.

For the detection stage, the researchers fine-tuned two classic convolutional neural network architectures, ResNet-18 and VGG-16, on the widely used Mendeley OCT dataset, which contains roughly 108,000 labeled images across four categories: normal, choroidal neovascularization, diabetic macular edema, and drusen. The team remapped these into three clinically meaningful classes—wet AMD, no AMD, and dry AMD—and addressed a significant class imbalance in the training data, which held about 51,000 normal images but only around 8,600 drusen cases, using augmentation techniques such as horizontal flipping and rotation. Both networks were initialized with ImageNet pre-trained weights and adapted to the retinal task.

The results were striking. On a balanced test set of 1,000 images, ResNet-18 achieved 98 percent accuracy, edging out VGG-16 at 96 percent, with ResNet-18 showing particular strength in detecting drusen—likely thanks to its residual connections, which help gradients flow through deep networks and preserve sensitivity to subtle pathological patterns. Notably, the researchers also benchmarked a Vision Transformer, the architecture behind many recent AI breakthroughs, and found it reached only 95 percent accuracy. The convolutional networks, it turns out, still hold their own on modest-sized medical datasets, where transformers’ lack of built-in spatial biases can leave them at a disadvantage.

The second stage is where the study breaks new ground. With no severity labels available, the team extracted 512-dimensional feature vectors from ResNet-18’s penultimate layer for all 8,616 drusen images, then projected them into a lower-dimensional space using UMAP, a technique chosen over linear methods like PCA because retinal features lie on a non-linear manifold. K-means clustering with three groups—chosen to mirror the Beckman Classification’s clinically defined early, intermediate, and advanced stages—was then applied. The clustering achieved a silhouette score of 0.49, indicating moderate but meaningful separation, exactly what one would expect given that drusen progression is continuous rather than sharply compartmentalized.

Crucially, the clusters were not just statistical artifacts. The researchers independently measured drusen area and equivalent diameter from the raw OCT scans using a classical segmentation pipeline, and found a statistically significant monotonic increase in drusen burden across the three clusters, confirmed by a Kruskal-Wallis test with a p-value below 0.001. Mean drusen area rose from about 36,400 pixels in the earliest cluster to over 46,300 pixels in the most advanced one. Interestingly, the separation between early and intermediate stages was modest—a 5.8 percent difference in mean area—while the gap between intermediate and advanced disease was far more pronounced at 20.3 percent, suggesting the algorithm is better at flagging advanced disease than at drawing fine distinctions in the earliest phases.

The team also stress-tested their models against domain shift, the phenomenon where models trained on data from one scanner or clinic falter on data from another. Without adaptation, the Mendeley-trained models collapsed to roughly 44 percent accuracy on the Duke OCT dataset. But after transfer learning, performance rebounded to between 82 and 84 percent, demonstrating that the framework can adapt to different imaging environments—a critical property for any tool hoping to leave the laboratory. The authors candidly note a limitation in their evaluation: because the Duke data was split at the slice level rather than the patient level, the reported cross-dataset figures may be somewhat optimistic.

The researchers are careful to position the system as a decision-support tool rather than a replacement for ophthalmologists, and they acknowledge that the unsupervised severity grading remains exploratory until validated against expert annotations or a ground-truth dataset such as AREDS. Still, the combination of high detection accuracy, clinically anchored clustering, and demonstrated cross-dataset adaptability points toward a scalable future in which AI triages the growing flood of retinal scans, flags the patients who need attention most urgently, and helps preserve sight for millions before the disease advances beyond reach.

Subject of Research: Machine learning-based detection and unsupervised severity grading of age-related macular degeneration using retinal OCT images

Article Title: Detection and exploratory severity grading of age-related macular degeneration using machine learning

Article References: Soobhany, A. R., & Thafheem, M. (2026). Detection and exploratory severity grading of age-related macular degeneration using machine learning. Frontiers in Imaging, 5, Article 1888372. https://doi.org/10.3389/fimag.2026.1888372

Image Credits: AI Generated

DOI: 10.3389/fimag.2026.1888372

Keywords: age-related macular degeneration, optical coherence tomography, deep learning, ResNet, VGG-16, unsupervised clustering, drusen, severity grading, transfer learning, UMAP, K-means, retinal imaging

The core problem pyMOODS addresses is one that optimization researchers have long acknowledged but rarely solved. Multi-objective optimization does not yield one optimal design; it yields an ensemble of Pareto-optimal candidates, each representing a different compromise among objectives such as cost, generation capacity, and environmental impact. Choosing among them is a human judgment call, informed by policy goals and stakeholder priorities. Yet the tools available for this post-optimization stage are fragmented. Algorithmic toolkits like pymoo, PlatEMO, and Borg generate Pareto fronts but offer little for the decision-maker. Multi-criteria decision analysis libraries such as pymcdm and pyDecision rank alternatives once objectives are fixed, but expose no exploratory interface. Visualization tools like PAVED and Parasol support interactive exploration but stop short of preference-weighted ranking. What has been missing, the PNNL team argues, is a single interactive platform that connects the visualization layer to a complete decision-support workflow on fronts that have already been computed.

pyMOODS fills that gap with a decoupled software architecture that separates heavy computation from interactive exploration. The frontend is a React and MUI application rendering linked interactive plots built with React Plotly, while a Python backend handles the analytics. User actions issue asynchronous requests to a Flask REST API, which mediates all communication between the two. Because interaction state lives in the frontend while analysis runs in the backend, computation never blocks the interface. A data access module loads, validates, and caches datasets held as CSV and JSON files, and a data processing module performs transformations and filtering before invoking the analytical routines. Notably, the dashboard evolved from early Plotly Dash prototypes into the current React-based implementation, gaining dynamic rendering of parameters, plots, and objective functions, along with better responsiveness under high-dimensional workloads.

One of the framework’s most elegant design decisions is its declarative input schema. A problem formulation is supplied as a single JSON configuration file that links to the data files and assigns each column to one of five categories: input parameters, hyperparameters, decision variables, objective functions, and control inputs. This means a single deployment can serve formulations whose objectives, decision variables, and hyperparameters differ in number and meaning, without any code changes. That flexibility matters because the framework targets real infrastructure applications, such as offshore wind co-design, MTDC-ESS system design, and load-shedding-constrained network expansion, where problem structures vary dramatically across use cases. It is a practical requirement that benchmark-oriented tools, typically evaluated on synthetic test problems like ZDT and DTLZ, simply do not address.

The analytical heart of pyMOODS combines two very different classes of methods. For structuring the solution space, the framework turns to unsupervised machine learning: a Uniform Manifold Approximation and Projection (UMAP) embedding projects the high-dimensional objective and decision space into two dimensions, and density-based HDBSCAN clustering groups similar solutions within that latent space. This gives users a navigable map of the entire front, where clusters of comparable candidates are visible at a glance and outliers stand out immediately. Users can also plot any pair of objectives or decision variables directly, or restrict the working set to solutions sharing a chosen hyperparameter configuration, such as a particular technology option.

For trade-off interpretation, pyMOODS takes a deliberately different path: a deterministic, parameter-free rank analysis. Every solution is ranked on each objective, and a solution’s generalizability is defined as its worst rank across all objectives. A solution is said to specialize in an objective when its rank there is strictly better than that of every more generalizable solution. Because this characterization derives entirely from the rank matrix, it introduces no trained model, no tuned threshold, and no random seed; identical inputs always produce identical generalizer and specializer assignments. This answers a question that ranking alone cannot: not merely which solutions score well under a given weighting, but which specific compromise each one represents. Solutions can additionally be ordered by an aggregate score reflecting user-specified objective weights, giving stakeholders a familiar mechanism for expressing their priorities, and the ranking is computed from the stored front rather than by re-solving, so stakeholders with different priorities can interrogate the same results interactively.

The framework’s workflow comes alive in its demonstration on the MoCoDo dataset for offshore wind farm planning with battery energy storage. The problem involves six objectives, including battery cost, cable material cost, day-ahead revenue, real-time revenue, and two reserve revenue streams, along with two decision variables: cable capacity and battery rated power. A conventional approach would aggregate all revenue streams into one total and all costs into another, reducing the problem to a simple two-way trade-off, but that aggregation obscures how individual solutions perform across specific revenue mechanisms. pyMOODS instead guides the user through four linked views: a scatter plot offering a first overview of the solution space, a ranking table for selecting a candidate, a parallel coordinates plot distinguishing generalizers from specializers across all six objectives, and beeswarm plots showing the distribution of all candidates per objective and decision variable. Selection propagates across views, so a candidate can be examined from several perspectives without losing context.

Performance testing suggests the architecture delivers on its responsiveness promise. On a local deployment running on an Apple M2 Pro with 16 GB of RAM, the team profiled three representative user tasks across two built-in formulations, Cameo_datacenter and MoCoDo_v2, repeating each measurement three times. Formulation initialization took roughly five milliseconds in both cases, while full dashboard load sequences averaged 3.838 seconds for the larger formulation and 0.475 seconds for the smaller one, with filter and selection interactions showing nearly identical latencies. The release also includes automated frontend tests using Playwright, which verify that dashboard components load correctly, the solution table sorts properly, and the generalizer-specializer controls update results, using simulated backend responses to validate the interface without requiring the full backend to run.

A capability comparison against representative open-source tools underscores what makes pyMOODS distinctive. Front-generation frameworks and metaheuristic platforms are not designed to be driven by a decision-maker; multi-criteria decision analysis libraries rank alternatives but expose no exploratory interface; and interactive visualization tools support exploration but stop short of preference-weighted ranking. In the published comparison, pyMOODS is the only tool that simultaneously offers an interactive user interface, operates on precomputed fronts from any solver, supports user-weighted ranking, provides automatic trade-off characterization, and accepts schema-agnostic problem formulations. The authors are candid about the release’s limitations, however: performance has been characterized on only two formulations on a single machine, rank-and-weight aggregation cannot recover non-convex regions of a Pareto front, and no usability evaluation has yet been conducted. Planned features include additional MCDM methods for comparison and an AI chatbot assistant.

The implications reach well beyond software engineering. As power systems worldwide absorb rising penetrations of onsite energy sources and face intensifying extreme weather, the co-optimization of generation, transmission, and storage investments is becoming one of the defining computational challenges of the energy transition. Tools like pyMOODS aim to ensure that the enormous computational effort poured into optimizing these systems does not end in an unreadable list of trade-offs. Released under the BSD 3-Clause License and archived on GitHub and Zenodo, version v0.0.3 puts a complete post-optimization decision-support pipeline into the hands of the planners, engineers, and researchers who must ultimately decide which compromises our future infrastructure will embody. Whether it changes real-world decisions, the authors acknowledge, is a question that future work, and future users, will have to answer.

Subject of Research: An open-source visual analytics framework for multi-objective decision support in large-scale infrastructure planning

Article Title: pyMOODS: An open-source framework for multi-objective decision support in large-scale infrastructure planning

Article References: Mattoo, P., Pham, J. N., Jain, M., Arendt, D., Avila, P., Ramachandran, T., Adetola, V., Yun, J. Y., & Wenskovitch, J. (2026). pyMOODS: An open-source framework for multi-objective decision support in large-scale infrastructure planning. SoftwareX, 36, Article 103095. https://doi.org/10.1016/j.softx.2026.103095

Image Credits: AI Generated

DOI: 10.1016/j.softx.2026.103095

Keywords: pyMOODS, multi-objective optimization, Pareto front, visual analytics, infrastructure planning, offshore wind, UMAP, HDBSCAN, decision support systems, open-source software, energy systems, Pacific Northwest National Laboratory

The problem scDEAN targets is a familiar one in computational biology. Single-cell RNA-seq data are high-dimensional, with expression matrices dominated by zeros because most genes are not detected in most cells. This sparsity, combined with dropout events and technical noise, obscures the genuine biological variability that distinguishes a T cell from a fibroblast. Existing clustering pipelines tend to solve one problem at a time: some methods preserve gene expression variability well but lose the geometric relationships between cells, while graph-based approaches capture cellular topology but can distort the underlying expression signal. Methods that try to handle both often rely on computationally costly attention mechanisms or elaborate batch-correction procedures, placing them out of reach for laboratories without serious computing resources. The Dhanbad duo set out to show that a carefully designed, parameter-efficient architecture could deliver competitive accuracy at a fraction of the cost.

At the heart of scDEAN lies a dual-encoder architecture, a design that processes the same dataset through two parallel pathways and then merges their outputs. The first pathway is a negative binomial-based autoencoder, a neural network trained to compress and reconstruct the raw count matrix. The choice of the negative binomial distribution matters: it is a statistical model that naturally accommodates the overdispersed, zero-inflated character of sequencing counts, so the encoder learns representations that respect the true generative process of the data rather than treating counts as continuous Gaussian measurements. The second pathway is a graph autoencoder built on a UMAP-derived fuzzy graph. UMAP, a manifold-learning technique widely used for visualizing single-cell data, is repurposed here to construct a weighted neighborhood graph that encodes fine-grained cellular geometry, and the graph autoencoder then learns embeddings that preserve those similarity relationships.

The fusion of these two representations is where scDEAN departs most visibly from its predecessors. Instead of multi-head attention, the framework employs a softmax-gated mechanism that dynamically weights the contribution of each encoder’s output. The gate is computed from the representations themselves, so the model can lean more heavily on expression information for some datasets and more heavily on graph structure for others, all with minimal parameter overhead. Because the gating weights are explicit, the fusion is also interpretable: researchers can inspect how much each information channel contributed to the final embedding. In an era when attention blocks are often bolted onto architectures by default, the authors’ argument is that a simple, well-placed gate can achieve similar adaptivity at a small fraction of the computational price.

A second design decision addresses a chronic tension in deep clustering. Many frameworks couple representation learning and clustering into a single end-to-end objective, but the two goals can pull the network in conflicting directions: the reconstruction loss wants embeddings that faithfully encode the data, while the clustering loss wants embeddings that form tight, well-separated groups. scDEAN decouples the clustering module from the representation learning module, allowing each to optimize its own objective without interference. Within the clustering module, the task is formulated as a differentiable centroid optimization problem with explicitly learnable centroids. Rather than repeatedly reinitializing K-means, a common and unstable practice, the model learns fuzzy membership assignments that let each cell belong partially to multiple clusters, a natural fit for biology where cell states exist on continua rather than in discrete boxes.

To sharpen the resulting partition, the authors add a distance-based regularizer that encourages compact structure within clusters and clear separation between them. The combination of learnable centroids, fuzzy memberships, and geometric regularization means the clustering stage converges smoothly and reproducibly, avoiding the sensitivity to random initialization that plagues many pipelines. The overall effect is a framework whose components are individually simple, a count-aware autoencoder, a graph autoencoder, a softmax gate, and a differentiable clustering head, but whose integration addresses the joint preservation of expression variability and cellular topology that has eluded many heavier alternatives.

The empirical evaluation spans a broad collection of public scRNA-seq datasets, each capturing a different biological context and technical platform. These include human preimplantation embryo and embryonic stem cell datasets from Yan and Petropoulos and their colleagues, a human liver bud organoid dataset from Camp and co-workers, droplet-based embryonic stem cell data from Klein and colleagues, mouse cortex and hippocampus cells from Zeisel and collaborators, peripheral blood mononuclear cells from 10x Genomics, human and mouse pancreas cells from Baron and colleagues, and an Alzheimer’s disease entorhinal cortex atlas from Grubman and team. Across these benchmarks, scDEAN achieved competitive performance relative to existing scRNA-seq clustering methods, with improvements observed on several datasets and the comparisons supported by formal statistical analyses following established protocols for evaluating classifiers across multiple data collections.

Crucially, the authors did not stop at clustering metrics such as agreement with known labels. To test whether the algorithm recovers biology rather than artifacts, they extracted the top 50 saliency-derived genes, genes identified as most influential to the model’s representations using saliency map techniques, from a head and neck squamous cell carcinoma dataset originally published by Puram and colleagues. Pathway enrichment analysis of these genes revealed enrichment of functionally relevant biological pathways, suggesting that scDEAN’s clusters are organized around genuine molecular programs. The team then pushed the validation into the clinical domain: using bulk RNA-seq expression and survival data from the TCGA-HNSC cohort obtained through the UCSC Xena GDC hub, they performed survival analysis that supported the clinical relevance of the saliency-derived gene set, linking the model’s unsupervised discoveries to patient outcomes in head and neck cancer.

The biological signals uncovered in the validation are consistent with known cancer biology. Genes highlighted by the framework connect to themes such as the behavior of cancer-associated fibroblasts, the stromal cells that shape tumor progression and immunotherapy resistance, and families of proteins implicated in metastasis and prognosis in head and neck squamous cell carcinoma. While the authors are careful to frame these analyses as validation rather than novel clinical claims, the exercise demonstrates the practical payoff of a clustering method that preserves biologically meaningful structure: the groups it finds can be interrogated directly for pathway activity and prognostic value, closing the loop between an unsupervised machine learning output and downstream biological interpretation.

Accessibility and reproducibility round out the package. The source code for scDEAN is publicly available on GitHub, and the exact processed versions of all datasets used in the study are deposited on Figshare, with the original data remaining available through GEO, ArrayExpress, and the 10x Genomics portal. For a field in which methodological papers sometimes outpace their own reproducibility, this openness lowers the barrier for other groups to benchmark, extend, or deploy the framework. The work, conducted at the Department of Mathematics and Computing at IIT (ISM) Dhanbad without dedicated external funding, is a reminder that progress in computational biology does not always require bigger models. Sometimes it requires a sharper question, in this case, how to fuse expression and geometry without excess machinery, and an architecture disciplined enough to answer it efficiently. If the competitive results hold up as laboratories adopt the tool, scDEAN may well become a reference point for lightweight, interpretable deep clustering in single-cell genomics.

Subject of Research: A lightweight adaptive deep learning model for clustering single-cell RNA sequencing data

Article Title: scDEAN: a lightweight adaptive fusion model for clustering scRNA-seq data

Article References: Chatterjee, S., & Bose, R. (2026). scDEAN: a lightweight adaptive fusion model for clustering scRNA-seq data. Applied Intelligence, 56(15), Article 455. https://doi.org/10.1007/s10489-026-07502-9

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07502-9

Keywords: scRNA-seq, clustering, deep learning, autoencoders, graph neural networks, dimensionality reduction, single-cell transcriptomics, bioinformatics, UMAP, head and neck cancer, survival analysis, machine learning