The research, conducted by Hrushikesh Sanap, an independent researcher based in Chhatrapati Sambhajinagar, India, evaluates four architecturally diverse deep networks alongside two ensemble strategies across four datasets drawn from the MedMNIST v2 benchmark collection. The individual models span the modern computer vision landscape: ConvNeXt-Base, a convolutional network redesigned with contemporary training recipes; Vision Transformer Base, an attention-based architecture that treats image patches as sequences; EfficientNetV2-M, a compact and efficient convolutional family; and InceptionResNetV2, an older but still competitive hybrid of inception modules and residual connections. All four were fine-tuned from ImageNet pre-trained weights at a resolution of 224 by 224 pixels, a deliberately realistic setting for clinical pipelines where computational budgets matter.

The four tasks were chosen to cover a spread of modalities, class counts and imbalance regimes. BloodMNIST requires eight-class classification of blood cell microscopy images and is close to saturation in the literature. BreastMNIST is a binary malignancy detection problem on breast ultrasound, and notably the smallest dataset in the study. DermaMNIST involves seven-class dermatoscopic lesion classification with pronounced class imbalance, and OrganAMNIST asks models to identify one of eleven abdominal organs on computed tomography slices. Every trained configuration exceeded the previously reported benchmark accuracy on all four datasets, with margins ranging from a fraction of a percentage point on the near-saturated blood cell task to roughly fifteen percentage points on the dermatoscopy data.

Two ensemble strategies were then layered on top. Soft Voting simply averages the predicted probabilities of the member networks, a technique with roots stretching back to foundational work on neural network ensembles in 1990. Rigorous Stacking, by contrast, trains a meta-learner on the outputs of the base models, following the stacked generalization framework introduced by David Wolpert in 1992. The ensembles matched or exceeded the strongest individual model in almost every case, with a single exception: Rigorous Stacking on the small BreastMNIST set, where the meta-learner presumably lacked enough data to learn reliable combination weights.

The central and most striking finding is that accuracy and calibration are dissociated. Calibration, measured here through reliability diagrams and the expected calibration error, asks whether a model’s stated confidence matches its empirical hit rate: of all the cases a model labels with 80 percent confidence, roughly 80 percent should actually be correct. Across the experiments, the most accurate configuration was rarely the best calibrated. Soft Voting attained the lower expected calibration error on all four datasets and the higher accuracy on three of them, while the accuracy advantage of Rigorous Stacking was confined to the most imbalanced dataset, DermaMNIST, and even there it came at the cost of weaker, less usable probabilities. Overconfidence, a well-documented pathology of modern neural networks, proved pervasive among the individual models and was reduced, but not eliminated, by ensembling.

Predictive uncertainty was quantified through the Shannon entropy of each prediction’s probability distribution, comparing the entropy distributions of correct against incorrect answers. Throughout the experiments, incorrect predictions sat at higher entropy than correct ones, which is the essential precondition for confidence-based triage, the practice of routing uncertain cases to human experts while letting confident, correct predictions pass through automatically. But the study adds a crucial caveat: a strong aggregate calibration score can still mask a collapsed, non-informative per-prediction uncertainty distribution. In other words, a model can look well calibrated on average while being nearly useless at telling clinicians which individual cases deserve scrutiny. Soft Voting held the most consistent separation between the entropy distributions of correct and incorrect predictions, reinforcing its case as the dependable default.

Interpretability analyses using gradient-weighted class activation mapping for the convolutional models and attention-based methods for the transformer complemented the quantitative results, and the residual errors clustered at clinically recognized decision boundaries. In dermatoscopy, the melanoma-to-nevus distinction remained the dominant failure mode, and even the largest accuracy gain left roughly a quarter of melanomas assigned to the benign nevus category. In breast ultrasound, the dangerous errors were malignant false negatives, the cases where a cancer is waved through as benign. In abdominal CT, the models struggled with kidney laterality, a reminder that spatial reasoning failures can persist even when overall organ identification accuracy looks impressive.

The practical implications are pointed. The study argues that post-hoc calibration should be treated as a prerequisite for safe deployment rather than an optional refinement, and that ensemble methods for medical imaging should be evaluated on calibration and uncertainty structure as much as on discrimination accuracy. For hospitals and regulators weighing which systems to trust, this reframes the leaderboard: a marginally less accurate but better calibrated ensemble that knows when it does not know may save more lives than a sharper classifier that is confidently wrong at the worst moments. The finding that simple probability averaging outperformed a learned stacking layer on calibration, while costing almost nothing in accuracy, is a rare case of the cheaper and simpler option also being the safer one.

The author is careful to delineate the limits of the evidence. All results derive from single-source benchmark data at 224 by 224 resolution, from a single training run under one fixed random seed, so run-to-run variance and dataset shift remain untested. Higher-resolution inputs and prospective external validation on real clinical data are flagged as the necessary next steps toward deployment. The datasets themselves are publicly available through the MedMNIST v2 collection, and the model prediction outputs, ensemble implementation and evaluation code have been released in an open GitHub repository, allowing independent verification of the per-class, entropy and reliability values underlying the analysis. For a field racing to put diagnostic AI in front of patients, the message is uncomfortable but timely: before asking how often a model is right, ask whether it knows the difference between knowing and guessing.

Subject of Research: Calibration and uncertainty quantification in multi-architecture deep learning ensembles for medical image classification on MedMNIST benchmarks

Article Title: Beyond accuracy: calibration and uncertainty in multi-architecture ensembles for MedMNIST classification

Article References: Sanap, H. (2026). Beyond accuracy: calibration and uncertainty in multi-architecture ensembles for MedMNIST classification. Neural Computing and Applications, 38(19), Article 776. https://doi.org/10.1007/s00521-026-12521-1

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12521-1

Keywords: medical image classification, ensemble learning, model calibration, uncertainty quantification, MedMNIST, expected calibration error, vision transformers, ConvNeXt, soft voting, stacked generalization, predictive entropy, clinical decision support

The core challenge the researchers set out to solve is a fundamental blind spot in existing machine learning approaches to mineral prospectivity modeling. Convolutional neural networks, the workhorses of modern image recognition, excel at spotting local patterns in voxelized 3D geological models, such as the geometry of a fault zone or the contact between an igneous intrusion and its host rock. But convolutional operations have inherently limited receptive fields, meaning they struggle to represent broader spatial relationships: the reach of a regional fault system, the alignment of intrusive bodies, or the distributed structural architecture that channels hydrothermal fluids across kilometers of crust. Graph neural networks can capture exactly those long-range relationships, yet they sacrifice detailed volumetric information. Neither approach alone, the authors argue, can fully characterize the multi-scale structure of a complex mineral system.

Their solution marries the two. The first branch of the hybrid model is a ResNet3D backbone enhanced with squeeze-and-excitation modules, which processes 7 by 7 by 7 voxel patches centered on each sampling location. The squeeze-and-excitation mechanism performs global average pooling over 3D feature maps and then recalibrates channel-wise responses through a small two-layer network, allowing the model to emphasize geological attributes most strongly associated with mineralization, such as Triassic host formations, diorite intrusions, and fault-related zones. The second branch is built on GATv2, a modern graph attention network, in which every sampling location becomes a node connected to its spatially nearest neighbors in a dynamically constructed K-nearest-neighbor graph. Attention weights, modulated by radial-basis distance encoding, let each node learn how much to trust information from its neighbors, effectively encoding coordinate-based spatial context that the convolutional branch cannot see.

Perhaps the most conceptually elegant component is the fusion mechanism that stitches these two branches together. Rather than averaging the two representations with fixed weights, the researchers introduce an entropy-guided adaptive gated fusion module. Each branch produces a probability prediction, and from that probability the team computes predictive entropy, a measure of how uncertain or ambiguous that branch is about a given location. Low entropy signals a confident branch; high entropy signals hesitation. The fusion gate reads these entropy values alongside measures of feature complementarity and dynamically adjusts the weighting alpha between the CNN embedding and the graph embedding, location by location. Where geology is locally complex and one branch falters, the other takes the lead. The authors are careful to note that this is a reliability-aware weighting scheme, not a full Bayesian uncertainty quantification framework, but it gives the model a self-correcting instinct that simple concatenation lacks.

Building the evidence base for such a model was a formidable undertaking in itself. The team integrated 1:50,000-scale geological maps, 86 mine-scale and regional cross sections, data from 489 boreholes, 26 audio-magnetotelluric interpreted profiles, and historical exploration reports. From these they constructed a 3D geological framework of the Anqing area, discretized into a voxel grid with 50-meter cubes, a resolution chosen as a compromise between geological fidelity and computational feasibility. The full prediction domain contained roughly 14.5 million valid voxels. For supervised training, the researchers extracted 4,253 voxels at known mineralized locations as positive samples and carefully selected 4,253 candidate negatives. Crucially, they avoided the easy trap of comparing ore against geologically irrelevant background: negative candidates were stratified by distance to known mineralization, from within 250 meters out to beyond 500 meters, and matched by geological signature, forcing the model to learn genuinely discriminative near-ore patterns. The authors candidly acknowledge that candidate negatives cannot be considered absolutely barren, since undiscovered mineralization may lurk within them.

When the full hybrid model was put to the test, the numbers spoke loudly. On the validation set it achieved an area under the precision-recall curve of 0.942 and an area under the receiver operating characteristic curve of 0.964, with roughly 93 percent accuracy and an F1 score of about 0.92. The comparisons were unforgiving: the GAT-only model managed an AUROC of 0.918, the CNN-only variant 0.844, and classical machine learning baselines such as logistic regression, random forests, and histogram gradient boosting languished with AUPRC values of only 0.59 to 0.61. The fusion model also proved robust across decision thresholds, holding an F1 near 0.92 across a wide range before degrading at extreme values, and it converged faster and more smoothly than either single branch. Embedding visualizations using principal component analysis and t-SNE showed the hybrid model producing far cleaner separation between mineralized and barren classes than the CNN alone, evidence that the graph module was genuinely integrating spatial adjacency information.

Turning a cloud of raw probabilities into something an exploration geologist can actually use required a further layer of engineering. The full-domain probability volume was smoothed, thresholded conservatively at 0.76, cleaned with binary morphological operations, and segmented using 3D connected-component analysis, volume filtering, vertical-continuity filtering, and geological-association screening. This pipeline suppressed isolated, overconfident voxels and retained spatially coherent bodies. The final output comprised five distal prediction targets, T1 through T5, none overlapping known mineralization, plus a sixth reference target supported by known mineralization that demonstrated the workflow could recover familiar ore-controlling geology. Maximum probabilities of the targets ranged from 0.916 to 0.978, and target-level mean entropy values between 0.343 and 0.431 provided a relative reliability index for ranking them.

The geological stories behind the individual targets are telling. Target T1, the largest, sits several hundred meters from the reconstructed Triassic host body and within a kilometer of diorite, but roughly 5.49 kilometers from known mineralization, making it a distal hypothesis rather than a near-mine extension. Target T2, the most distant at about 12.62 kilometers from known ore, overlaps only a sliver of Triassic-related rock and is flagged for caution given its weak intrusive association. Targets T3 and T5, by contrast, show stronger direct geological support, overlapping or abutting both the Triassic host rock and the diorite body, precisely the intrusive-host interaction that generates skarn copper mineralization. Target T4 overlaps the Triassic body within about 187 meters of diorite. The team stresses that all five remain predictive exploration hypotheses requiring independent geological, geophysical, and drilling validation before they can be called discoveries.

The study’s honesty about its own limitations may prove as influential as its results. The authors explicitly warn that validation metrics derive from a random-stratified sample-level split that cannot fully eliminate spatial autocorrelation, that the batch-wise KNN graph is a local approximation rather than a fixed full-region graph, and that the entropy-based fusion should not be read as full Bayesian uncertainty. Predictive entropy maps, displayed in plan view and in vertical cross sections, are offered as relative reliability guides for target ranking, not absolute confidence statements. In a field where high validation scores are too often mistaken for exploration certainty, this leakage-aware, uncertainty-conscious workflow sets a standard. If the hybrid voxel-and-graph vision of the subsurface holds up under the drill bit, the era of AI-guided mineral discovery in deeply concealed terrains may have quietly begun beneath the rice paddies of Anhui Province.

Subject of Research: Hybrid 3D CNN and graph attention deep learning for mineral prospectivity modeling in the Anqing skarn copper district

Article Title: A Hybrid 3D CNN-GAT Framework with Entropy-Guided Adaptive Fusion for 3D Mineral Prospectivity Modeling

Article References: Chen, C., Zhang, M., Wang, X., Wang, L., & Li, X. (2026). A Hybrid 3D CNN-GAT Framework with Entropy-Guided Adaptive Fusion for 3D Mineral Prospectivity Modeling. Natural Resources Research. https://doi.org/10.1007/s11053-026-10780-2

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10780-2

Keywords: mineral prospectivity mapping, deep learning, 3D geological modeling, graph attention network, convolutional neural network, predictive entropy, skarn copper deposit, Anqing, exploration targeting, uncertainty quantification, ResNet3D, Yangtze River Metallogenic Belt