Mpox-Trinet 0.9.0

Tri-Net v2 — reproducible deep-learning framework for Mpox skin-lesion diagnosis

Tri-Net v2

A Reproducible Deep-Learning Framework for Multi-Class Skin-Lesion and Symptom-Based Monkeypox (Mpox) Diagnosis

Official implementation of the Tri-Net study — upgraded with modern backbones, learned feature fusion, and a fully reproducible evaluation framework.

Original Publication • Abstract • What's New in v2 • Results • Integrity Audit • Quick Start • Model Zoo • Docs • Citation

_{14-class confusion matrix — leakage-free split, real test set, regenerated by trinet evaluate --champion}

Original Publication

Tri-Net: Unified Deep Learning for Skin Lesion and Symptom-Based Monkeypox Detection Sudharsan S¹ · Prabu Selvam²ᵗ · Nirmala Veeramani³ᵗ · Kiran Kumar B⁴ · Nikola Ivković⁵ · Korhan Cengiz⁶,⁷ ¹²⁴ School of Computing, SRM Institute of Engineering and Technology, Tiruchirappalli Campus, India · ³ School of Computing, SASTRA Deemed University, India · ⁵ Faculty of Organization and Informatics, University of Zagreb, Croatia · ⁶ College of Computing and Intelligent Systems, University of Khorfakkan, UAE · ⁷ Department of Electrical Engineering, Biruni University, Istanbul ᵗ Corresponding authors

Status: currently under peer review at Scientific Reports (Springer Nature, Nature Portfolio). Both reviewers have recommended the manuscript for acceptance; outstanding editorial requirements are (a) a formal Code Availability statement pointing to a DOI-assigning archive of this repository, and (b) provenance/permission documentation for a small number of figures. This repository is the code artifact referenced by that statement — see Code Availability below.

This repo is not the paper's original codebase — that codebase was, at the time this framework was built, largely unimplemented against the paper's own claims (see the Integrity Audit). Tri-Net v2 is a from-scratch, tested, reproducible rebuild that keeps the paper's core idea — a Tri-Net ensemble plus a symptom-based classifier — and makes every number in it independently verifiable. The original scripts are preserved unmodified under archive/ for provenance and direct comparison.

Abstract

The 2022–2024 re-emergence of Monkeypox (Mpox) renewed interest in rapid, non-invasive computer-aided diagnosis. Tri-Net proposes a unified deep-learning approach that couples skin-lesion image analysis with symptom-based prediction. Tri-Net v2 re-engineers the study into an installable, tested, and fully reproducible framework and modernises it end to end. It evaluates a challenging 14-class dermatological setting under a strictly leakage-free train/validation/test split, benchmarks five convolutional backbones (including current-generation ConvNeXt and EfficientNetV2) as frozen feature extractors, and introduces a learned feature-fusion ablation across five strategies (mean, PSO-weighted, concatenation-MLP, gated attention, transformer encoder). The best configuration attains 77.2% top-1 accuracy and 97.0% macro-AUC on the fine-grained 14-class task, while the clinically-actionable binary Mpox-vs-rest screening formulation reaches 98.4% accuracy and 99.6% AUC with reported sensitivity, specificity, and Wilson confidence intervals. The framework ships a complete evaluation suite (5-fold cross-validation, McNemar's test, Cohen's κ, ensemble diversity, Grad-CAM), a unified CLI, CI, Docker, and a model zoo. Every figure and number is regenerated from code — nothing is hand-typed.

What's New in v2

• Modern backbones — ConvNeXt-Tiny (the strongest single backbone, 69.3%) and EfficientNetV2-S added alongside the original EfficientNetB4 / InceptionResNetV2 / DenseNet201.
• Learned feature fusion — a five-way ablation (Mean, PSO, Concat-MLP, Gated-Attention, Transformer); feature-level fusion is the champion.
• Rigorous 14-class benchmark — a leakage-free 80/10/10 split (original images only) that never lets augmented copies straddle train/val/test — see why this matters.
• Dual-task evaluation — fine-grained 14-class diagnosis reported separately from binary Mpox screening, each with appropriate metrics and confidence intervals.
• Full reproducibility — installable trinet package, unified CLI, unit tests, GitHub Actions CI, Docker, and a model zoo. Every number regenerates from committed code.
• Complete evaluation suite — confusion matrices, ROC, Cohen's κ, McNemar's test, 5-fold CV, ensemble diversity analysis, and Grad-CAM interpretability — none of which existed in the original repository.
• Checkpoint integrity — a save-time protocol verifies that serialized models reproduce identical logits across processes, so deployed weights behave exactly as evaluated.

Results

Task	Metric	Score
Binary Mpox screening (clinical task)	Accuracy / AUC	98.35% / 99.58%
14-class fine-grained diagnosis	Accuracy	77.23%

The 14-class result is on a demanding benchmark: 14 classes, leakage-free 80/10/10 split, ~3,000 real images, cross-validated — a substantially harder setting than the binary or low-cardinality tasks common in the Mpox imaging literature, and a smaller, non-inflated corpus compared to the original paper's split (see the audit below).

Headline progression (click to expand)

Configuration	Test accuracy
Baseline ensemble (three original backbones, Mean/PSO)	72.60%
+ ConvNeXt-Tiny backbone (Mean ensemble)	75.25%
+ Concat-MLP feature fusion (champion)	77.23%

Net gain from the v2 upgrades: +4.63 points. The ensemble's diversity-derived oracle ceiling is ~80%, so fusion captures most of the headroom available in this ensemble.

Single-backbone benchmark (frozen features)

Backbone	Accuracy	Macro-F1	AUC	κ
EfficientNetB4	58.4	63.3	95.0	54.2
DenseNet201	66.3	67.2	95.6	62.8
InceptionResNetV2	68.0	68.8	96.2	64.4
ConvNeXt-Tiny	69.3	70.3	96.5	65.9

Fusion-strategy ablation (best-3 backbones)

Strategy	Learnable	Params	Accuracy	Macro-F1	AUC	κ
Concat + MLP	yes	2.2 M	77.2	79.1	97.0	74.6
Mean	no	0	75.3	77.2	97.1	72.5
PSO	global	3	73.3	74.7	96.8	70.4
Transformer	yes	1.5 M	65.0	68.3	96.6	61.6
Gated Attention	yes	1.2 M	63.4	63.2	95.6	59.8

Feature-level fusion beats probability fusion; the highest-capacity variants overfit at this dataset size.

Cross-validation, ensemble diversity & per-class

5-fold CV (mean ± SD): EfficientNetB4 59.0 ± 2.6 · InceptionResNetV2 59.2 ± 1.4 · DenseNet201 65.5 ± 0.6 · Tri-Net 69.3 ± 0.8 (beats every base in all folds).

Diversity: pairwise disagreement 24.9%, Yule's Q 0.75, error-correlation ρ 0.43, double-fault 19.7% → variance reduction, oracle ceiling ≈ 80%.

Per-class split: pox/infectious classes 89.3% vs dermoscopy (skin-cancer) classes 73.2% — confusions concentrate on clinically hard pairs (Melanoma↔Nevus, BCC↔SCC).

Binary screening — full breakdown

Metric	Score	95% CI (Wilson)
Accuracy	98.4%	96.2 – 99.3
AUC	99.6%	—
Sensitivity	86.2%	69.4 – 94.5
Specificity	99.6%	98.0 – 99.9

Specificity is very high; sensitivity has a wider interval (few positive test cases). Threshold tuning toward sensitivity is a natural next step — see docs/benchmark.md.

Full tables and ablations: docs/benchmark.md.

Reproducibility & Integrity Audit

This is the core contribution of v2: a documented, evidence-based account of where the original paper's claims and its accompanying code diverged, verified directly against the manuscript and its original repository.

#	Finding	Evidence
1	The paper's PSO ensemble had no corresponding implementation in the original repository — the code simply concatenated backbone features. We built a real PSO optimizer; under our reproduction it underperforms a plain probability mean (73.3% vs 75.3%).	Paper §3.4 & Algorithm 1 claim PSO weights (0.42, 0.32, 0.26); original lesion.py contained no optimizer.
2	No trained 14-class checkpoint was ever saved in the original repository — the 97.86%/97.86%±0.29 figures reported across the paper's Tables 3, 6, 8, and 9 have no retrievable model artifact behind them.	Repository inspection at rebuild time.
3	The paper is internally inconsistent about which module scored 97.86%. The abstract attributes it to "the symptom-based module alone," while the body's Tables 3, 6, 8, and 9 all report the same figure for the image-based Tri-Net ensemble instead.	Direct comparison of Abstract vs. §4.3–4.5.
4	Two different confusion matrices for the same binary evaluation don't agree. Fig. 6 reports TP=905, FN=16, FP=13, TN=431 (n=1,365); the accompanying paragraph reports TP=920, FN=12, FP=20, TN=445 (n=1,397) — different totals, different every cell. A third figure elsewhere (§4.4) describes 3,710 test images with 3,598 correct / 111 incorrect, which matches neither.	Paper Fig. 6 vs. the paragraph immediately following it vs. §4.4.
5	Section 4 opens with a paragraph describing an unrelated study — a routing/optimization paper about energy-aware networking — rather than the Mpox experiments actually being reported, a clear copy-paste artifact from a different manuscript.	Paper §4, opening paragraph.
6	The claimed 97.86% is very likely inflated by augmentation leakage. The paper's 14-class split totals 29,322 images (23,458 train / 2,932 val / 2,932 test) built by heavy augmentation of a much smaller base — MSLD ships pre-augmented copies, and augmenting before splitting lets near-duplicate images straddle train/val/test. Our leakage-free split (original images only, augmentation applied at train time only) totals 3,032 images — roughly 1/10th the size — and yields 77.23%, not 97.86%.	Paper §3.1 vs. docs/datasets.md.
7	None of the paper's stated evaluation methodology existed in the original code — 5-fold CV, McNemar's test, Cohen's κ, ROC/AUC, Grad-CAM. We built the full suite from scratch; see docs/benchmark.md.	Repository inspection vs. paper §4.5–4.7.
8	The symptom-based classifier's headline number relied on a leaking feature. Dropping the sum/leakage-prone feature and retraining honestly yields ~64–70% across CNN, Logistic Regression, Random Forest, and XGBoost — far from any 97%+ figure.	docs/datasets.md, docs/benchmark.md.
9	Fine-tuning does not help at v2's data scale — unfreezing backbone layers on ~2.4k images overfits immediately and matches the frozen-ensemble baseline, contradicting the assumption that more training freedom is strictly better.	docs/benchmark.md.
10	Higher-capacity fusion (Transformer, gated attention) generalizes worse than simple Concat-MLP fusion at this dataset size — a statement about the data regime, not the architectures.	Fusion ablation table above.

None of this is a claim of bad faith — research code commonly lags a manuscript, and the reviewers' own comments note the revised manuscript now "adequately addresses" prior concerns. But reproducibility is the point of publishing quantitative results, and this repository exists to make every one of the numbers above independently checkable, permanently, from committed code. The original paper's scripts are preserved unmodified under archive/ for anyone who wants to verify these findings directly.

Methodology

Backbones & transfer learning. Five ImageNet-pretrained backbones are compared, each with its own required preprocessing. Each frozen backbone feeds a lightweight head (Dense→BatchNorm→Dense→Dropout→Softmax); features are cached once so training runs in seconds/epoch. Class imbalance is handled with balanced class weights.

Fusion strategies. Five ways to combine backbones on identical features: Mean (equal-weight probability average), PSO (a global weight vector optimised by Particle Swarm Optimization), Concat-MLP (concatenate features → MLP), Gated Attention (per-image attention over backbone tokens), and Transformer (self-attention encoder).

Symptom module. A compact CNN with Logistic-Regression / Random-Forest / XGBoost baselines on the structured symptom dataset, using a leakage-free feature pipeline.

Evaluation protocol. Accuracy, macro-P/R/F1, one-vs-rest macro-AUC, and Cohen's κ on the held-out test set; 5-fold CV on original features; McNemar's test; ensemble diversity; Grad-CAM; and Wilson CIs for binary screening. Full details: docs/architecture.md.

Architecture

datasets ──► features ──► base heads ──► ensemble ─┐
 (download,   (frozen      (per-backbone   (PSO /   ├─► evaluation
  prepare)     backbone     classifier)     mean)   │    (metrics, confusion,
               cache)                               │     ROC, Kappa, McNemar,
                                    fusion ─────────┘     diversity, Grad-CAM)
                                 (Concat-MLP / attention / transformer)

Backbones are frozen and cached to disk as feature vectors, so head training runs in seconds per epoch rather than hours.

Backbone	Pooled dim	Notes
EfficientNetB4	1792	Metal-compatible
InceptionResNetV2	1536	Metal-compatible
DenseNet201	1920	Metal-compatible
ConvNeXt-Tiny	768	best single backbone; CPU extraction on Apple Metal
EfficientNetV2-S	1280	CPU extraction on Apple Metal

Full details: docs/architecture.md.

Quick Start

git clone https://github.com/Sudharsanselvaraj/Synergistic-Deep-Learning-for-Monkeypox-Diagnosis.git
cd Synergistic-Deep-Learning-for-Monkeypox-Diagnosis
python -m venv .venv && source .venv/bin/activate
pip install -e .              # add ".[metal]" on Apple Silicon, ".[dev]" for tests/linting

Run the full pipeline:

trinet download        # fetch datasets (needs a Kaggle API token)
trinet prepare         # leakage-free 14-class split + clean symptom data
trinet features        # cache frozen-backbone features
trinet train           # train classification heads
trinet fusion          # fusion-strategy ablation (Concat-MLP is the champion)
trinet evaluate --champion   # figures + tables from the best model

Or the whole image pipeline in one shot: trinet benchmark. Full CLI reference: trinet --help.

Inference on a single image:

trinet predict lesion.jpg

from trinet.inference.predict import predict_image
print(predict_image("lesion.jpg"))

As a library:

from trinet.models.fusion import build_concat_mlp
from trinet.evaluation.metrics import compute_scores

Model Zoo

Model	Task	Accuracy	AUC	Weights
Tri-Net v2 (Concat-MLP fusion)	14-class	77.2	97.0	✅
ConvNeXt-Tiny (single backbone)	14-class	69.3	96.5	✅
Mean ensemble	14-class	75.3	97.1	—
PSO ensemble	14-class	73.3	96.8	✅
Tri-Net v2 (Mpox screening)	binary	98.4	99.6	✅

Full breakdown, params, and Macro-F1 / Kappa per model: docs/model_zoo.md. Pretrained weights are attached to the matching GitHub Release.

Documentation

Doc	Covers
docs/architecture.md	Models, fusion strategies, full pipeline
docs/model_zoo.md	Pretrained models, weights, usage
docs/datasets.md	Data sources, licensing, 14-class composition
docs/benchmark.md	Full results and ablations
docs/reproducibility.md	How to regenerate every number in this README
docs/roadmap.md	Project plan, milestones, Phase 2 goals

Code Availability

This repository is the code artifact referenced in the manuscript's Code Availability statement, as required by Scientific Reports' policy on custom computational tools. Recommended statement for the manuscript:

The custom code, trained model weights, and evaluation pipeline supporting the findings of this study are openly available at github.com/Sudharsanselvaraj/Synergistic-Deep-Learning-for-Monkeypox-Diagnosis, archived on Zenodo at DOI: pending — archive a tagged release via the GitHub–Zenodo integration to mint one.

To satisfy this: create a GitHub Release for the current tag, connect the repository to Zenodo (Zenodo → GitHub → toggle the repo on, then publish a release), and add the resulting DOI badge here and to the manuscript's Code Availability section.

Reproducibility

• Global seed 42 fixed across NumPy, TensorFlow, and scikit-learn.
• The 80/10/10 split is stratified, seeded, and deterministic — the same images always land in the same partition.
• Metrics are computed only on the held-out test set; cross-validation uses un-augmented features so augmented copies never straddle folds.
• Generated artifacts (outputs/, data/) are git-ignored by design — regenerate everything with the CLI, nothing paper-facing is hand-typed.

trinet download && trinet prepare && trinet features
trinet train && trinet fusion && trinet evaluate --champion
trinet cross-val && trinet diversity && trinet gradcam

Reference environment: Apple M-series, 16 GB, Python 3.11, TensorFlow 2.16.2. Note: tensorflow-metal hangs on EfficientNetV2-S and errors on ConvNeXt — extract those on CPU with trinet features --cpu. Full notes: docs/reproducibility.md.

Repository Structure

.
├── src/trinet/            installable package
│   ├── config.py          central configuration (paths, 14 classes, hyperparameters)
│   ├── cli.py             Typer command-line interface
│   ├── datasets/          dataset download + leakage-free preparation
│   ├── models/            backbones, fusion strategies, PSO ensemble
│   ├── training/          feature caching, head training, fine-tuning, symptom model
│   ├── evaluation/        metrics, figures, cross-validation, diversity, efficiency
│   ├── explainability/    Grad-CAM
│   ├── inference/         single-image prediction + model export
│   └── utils/             doctor / info diagnostics
├── experiments/           one-off studies (e.g. binary screening with Wilson CIs)
├── configs/               YAML experiment presets
├── docs/                  architecture, datasets, benchmark, reproducibility, model_zoo, roadmap
├── examples/              runnable usage examples (predict, batch, ONNX export, custom dataset)
├── scripts/               shell helpers (download, train, evaluate, benchmark)
├── tests/                 unit tests (metrics, models, fusion, dataset, checkpoint integrity)
├── docker/                Dockerfile + docker-compose
├── app/                   deployment stub (API / web / docker) — planned
├── archive/               original Tri-Net scripts, preserved unmodified
├── assets/                logo + figures used in docs and README
└── .github/               CI workflows + issue/PR templates

Limitations

• Public composite datasets limit demographic/acquisition diversity; the leakage-free test set is modest (303 images), so metrics carry non-trivial confidence intervals.
• No external clinical validation — results are experimental, not clinically certified.
• Image and symptom modules are trained on disjoint patient populations, so genuine multimodal (joint) fusion is future work pending a paired dataset.
• The dermoscopy sub-problem is inherently hard; ~85–90% is near the field-wide ceiling even with far larger datasets and metadata.

Research artifact, not a medical device. Do not use for clinical decision-making without regulatory approval and prospective validation. See SECURITY.md.

Roadmap

• Complete the backbone table (EfficientNetV2-S) and an efficiency/latency profile.
• A hierarchical classifier (pox-group vs dermoscopy-group → within-group heads), motivated by the strong group-separation result — the most promising route to lift fine-grained accuracy.
• Larger, higher-resolution dermoscopy data + patient metadata; self-supervised (DINOv2) features.
• A genuine multimodal model, contingent on a paired image+symptom dataset. See docs/roadmap.md.

Contributing

Contributions are welcome — see CONTRIBUTING.md for dev setup, coding guidelines, and PR expectations. The short version: reusable code lives in src/trinet/, experiment-specific code lives in experiments/, and every reported number must be regenerable from committed code on the held-out test set.

Citation

The original study is currently under revision at Scientific Reports (Springer Nature); a DOI will be added here once assigned. Until then, please cite the manuscript and this reproducibility framework as follows:

@article{selvaraj2026trinet,
  title   = {Tri-Net: Unified Deep Learning for Skin Lesion and Symptom-Based Monkeypox Detection},
  author  = {Selvaraj, Sudharsan and Selvam, Prabu and Veeramani, Nirmala and
             Kumar B, Kiran and Ivkovi{\'c}, Nikola and Cengiz, Korhan},
  journal = {Scientific Reports},
  publisher = {Springer Nature},
  year    = {2026},
  note    = {Under revision. DOI to be added upon acceptance.}
}

@software{trinet_v2_2026,
  author  = {Selvaraj, Sudharsan},
  title   = {Tri-Net v2: A Reproducible Deep-Learning Framework for Mpox Skin-Lesion Diagnosis},
  year    = {2026},
  url     = {https://github.com/Sudharsanselvaraj/Synergistic-Deep-Learning-for-Monkeypox-Diagnosis}
}

See CITATION.cff for the machine-readable, definitive entry.

Acknowledgments

Built on the MSLD v2 and ISIC skin-lesion datasets and the Kaggle Monkeypox symptom dataset — full sourcing and licensing in docs/datasets.md.

License

MIT