EndoXAI-RCT is an engineering and research prototype for explainable, multi-model clinical image review. It is not a medical device and is not intended for autonomous diagnosis or treatment.
|
The browser-accessible deployment demonstrates the EndoXAI-RCT software workflow in a live cloud environment. Deployment: Google Cloud Run
|
The Hugging Face Space provides the associated EndoXAI-RCT research implementation and demonstration environment. Domain: Dental panoramic imaging
|
EndoXAI-RCT is an explainable AI software architecture developed to investigate how multiple artificial-intelligence models can be orchestrated within a structured clinical image-review workflow.
The system uses panoramic dental radiographs as its application context, with particular emphasis on AI-assisted review related to Root Canal Treatment (RCT) assessment.
The project is deliberately broader than a conventional image-classification demonstration.
Instead of asking only:
“What does the model predict?”
EndoXAI-RCT asks:
“How should heterogeneous AI predictions become traceable evidence that can safely support a human reviewer?”
That change in perspective drives the architecture.
Artificial intelligence can detect patterns in medical images.
But prediction alone does not constitute a complete decision-support system.
A deployable review workflow has to address additional engineering questions:
- 🧠 Which AI model is responsible for which task?
- 🎯 Which model provides the principal task-specific prediction?
- 🧩 Which models provide advisory or contextual information?
- 🔀 What happens when different models produce different outputs?
- 🔎 What evidence supports a result?
- 🔥 Can the reviewer see interpretable evidence associated with model behaviour?
⚠️ What happens when an input is invalid or a model is unavailable?- 🧭 Which evidence is allowed to influence downstream routing?
- 🛡️ Where does software authority end?
- 👩⚕️ Who makes the final assessment?
EndoXAI-RCT is architected around these questions.
🩻 Image → 🛡️ Validation → 🧠 Multi-Model Inference → 🔎 Evidence → 🧭 Routing → 🔥 Explainability → 👩⚕️ Human Review
|
Models can be assigned explicit responsibilities rather than treating every output as equally authoritative. |
Different models can contribute complementary signals to the review process. |
Outputs are transformed into information intended to support traceable review. |
Visual explanations help expose image regions associated with model behaviour. |
|
Evidence can be handled according to its defined role rather than silently combining every prediction. |
Validation and software checks help prevent invalid inputs or unavailable components from being treated as valid evidence. |
The architecture separates AI-generated evidence from human decision authority. |
The workflow terminates in human interpretation rather than autonomous clinical action. |
The EndoXAI-RCT workflow begins before the AI model.
A panoramic image passes through a physical imaging process before it ever enters the software pipeline.
Understanding this boundary is important because the AI system operates on a digitally reconstructed image, not directly on the patient or imaging hardware.
Note: If the animated GIF has not yet been added to
assets/, this image will appear afterassets/EndoXAI_Physical_to_Digital_Animated.gifis committed to the repository.
| Stage | Process | Purpose |
|---|---|---|
| 01 | 👤 Patient positioning | Establish stable acquisition geometry |
| 02 | ☢️ X-ray acquisition | Capture projection information |
| 03 | 📡 Detector capture | Convert captured radiation information into digital signals |
| 04 | 🖼️ Image reconstruction | Construct the panoramic image |
| 05 | 📁 Export / transfer | Produce an image suitable for downstream software review |
| 06 | ☁️ Upload | Transfer the image into EndoXAI-RCT |
| 07 | 🛡️ Validation & preprocessing | Check and prepare the input |
| 08 | 🧠 Multi-model inference | Generate model-specific outputs |
| 09 | 🧭 Evidence & routing | Organize outputs according to defined roles |
| 10 | 🔥 Explainability | Generate interpretable visual evidence |
| 11 | 👩⚕️ Human review | Present results for final human interpretation |
EndoXAI-RCT separates the clinical-AI workflow into functional layers.
┌─────────────────────────────────────────────┐
│ HUMAN REVIEWER │
│ Final interpretation and judgement │
└──────────────────────▲──────────────────────┘
│
Reviewable evidence
│
┌──────────────────────┴──────────────────────┐
│ EXPLAINABILITY LAYER │
│ Visual evidence • findings • confidence │
└──────────────────────▲──────────────────────┘
│
┌──────────────────────┴──────────────────────┐
│ EVIDENCE & ROUTING LAYER │
│ Model role • evidence handling • synthesis │
└──────────────────────▲──────────────────────┘
│
┌──────────────────────┴──────────────────────┐
│ MULTI-MODEL AI LAYER │
│ Primary • Advisory • Contextual analysis │
└──────────────────────▲──────────────────────┘
│
┌──────────────────────┴──────────────────────┐
│ VALIDATION & PREPROCESSING │
│ Input checks • preparation • safety gates │
└──────────────────────▲──────────────────────┘
│
┌──────────────────────┴──────────────────────┐
│ IMAGE INGESTION │
│ Panoramic radiograph upload │
└─────────────────────────────────────────────┘
A central architectural concept in EndoXAI-RCT is that multiple models do not automatically have equal authority.
Different models can contribute different forms of information.
The architecture therefore distinguishes between model capability and model decision influence.
The primary model represents the model assigned the principal task-specific responsibility within the configured workflow.
Its output may contribute directly to the principal review result, subject to validation and evidence-handling logic.
Typical responsibilities include:
- principal task-specific prediction,
- confidence generation,
- structured output generation,
- evidence contribution,
- downstream explanation support.
Advisory models provide additional information that may help contextualize the primary result.
Their outputs should not silently become equivalent to the primary model simply because they are available.
Potential advisory contributions include:
- complementary classification,
- contextual detection,
- additional image-pattern information,
- secondary confidence signals,
- reviewer-supporting evidence.
Some model outputs may be useful primarily as context.
Contextual information can enrich the reviewer interface while remaining separated from the principal routing decision.
This distinction supports a core EndoXAI-RCT principle:
More models do not automatically mean more authority.
A deployable multi-model system needs more than a list of model files.
It needs an explicit understanding of what each model is permitted to do.
Conceptually, EndoXAI-RCT treats each model as a registered component with attributes such as:
Model
│
├── Identifier
├── Intended task
├── Role
│ ├── Primary
│ ├── Advisory
│ └── Contextual
│
├── Input requirements
├── Output type
├── Availability / health
├── Evidence contribution
├── Routing influence
└── Explainability support
This reduces the risk of accidentally treating heterogeneous model outputs as interchangeable.
EndoXAI-RCT is designed around the idea that AI output should become reviewable evidence, not merely a number displayed on a screen.
A useful review object may contain:
Prediction
+
Confidence
+
Model identity
+
Model role
+
Supporting visualization
+
Processing status
+
Evidence provenance
+
Review context
This provides a richer basis for interpretation than presenting only:
Prediction: Positive
Multi-model AI introduces an important problem:
EndoXAI-RCT addresses this at the architecture level by separating:
model execution
from
decision influence.
┌───────────────┐
│ Input Image │
└───────┬───────┘
│
Validation Gate
│
┌───────────────┼───────────────┐
│ │ │
▼ ▼ ▼
Primary Model Advisory Model Context Model
│ │ │
▼ ▼ ▼
Evidence A Evidence B Evidence C
│ │ │
└───────────────┼───────────────┘
│
Evidence Router
│
▼
Explainable Review Object
│
▼
Human Review
The purpose of routing is not to conceal disagreement.
It is to make the relationship between evidence and decision support explicit.
Explainability is an important component of the EndoXAI-RCT workflow.
Where supported by the model and implementation, visual explanation techniques can be used to identify image regions associated with model behaviour.
The repository includes Grad-CAM-related implementation components such as:
gradcam_resnet_lsl.py
Explainability is intended to help answer:
“What image regions were associated with this model output?”
rather than asserting:
“This visualization proves the clinical diagnosis.”
That distinction is essential.
Conceptually:
Panoramic Radiograph
│
▼
Neural Network
│
▼
Selected Feature Layer
│
▼
Gradient Information
│
▼
Activation Importance
│
▼
Heatmap
│
▼
Radiograph + Explanation Overlay
The resulting visualization can help a reviewer inspect whether model attention appears aligned with relevant image regions.
A heatmap is an explanation aid, not independent clinical evidence and not proof of causality.
A responsible AI pipeline should not assume that every uploaded file is suitable for analysis.
The validation stage can serve as a boundary between user input and AI inference.
Conceptually:
Upload
│
▼
File / image validation
│
├── Invalid ─────────► Reject / request correction
│
▼
Preprocessing
│
▼
Model-ready image
│
▼
Inference
Validation is therefore part of the AI safety architecture, not merely a user-interface feature.
A multi-model architecture must also consider partial failure.
Examples include:
- model artifact unavailable,
- unsupported input,
- preprocessing failure,
- inference exception,
- explanation generation failure,
- low-confidence result,
- incomplete evidence.
The desired engineering principle is:
Where possible, the system should distinguish between:
AVAILABLE
DEGRADED
UNAVAILABLE
FAILED
NOT_APPLICABLE
and expose relevant status to the review workflow.
EndoXAI-RCT is designed as decision support, not autonomous clinical decision-making.
AI SYSTEM
│
▼
Predictions & Evidence
│
▼
Explainable Results
│
▼
┌──────────────────┐
│ HUMAN REVIEWER │
└──────────────────┘
│
▼
Independent judgement
The software can:
- process images,
- execute AI models,
- calculate model outputs,
- organize evidence,
- generate visual explanations,
- present structured findings.
The software does not replace professional judgement.
The central research question behind EndoXAI-RCT is broader than whether a neural network can classify a dental radiograph.
The project investigates:
This shifts the engineering focus:
FROM
"How accurate is one model?"
TO
"How should multiple AI capabilities behave together
inside a reviewable software system?"
The panoramic dental-imaging workflow provides the application context for demonstrating this architecture.
01 Patient / Imaging Workflow
↓
02 Panoramic Image Generation
↓
03 Digital Export
↓
04 Secure Application Upload
↓
05 Validation
↓
06 Preprocessing
↓
07 Multi-Model Execution
↓
08 Model-Specific Outputs
↓
09 Evidence Construction
↓
10 Role-Aware Routing
↓
11 Explainability
↓
12 Human Review
The EndoXAI-RCT web application is intended to transform the underlying AI pipeline into a reviewer-oriented workflow.
The interface brings together:
- 🩻 image upload,
- 🛡️ preprocessing and validation,
- 🧠 AI inference,
- 📊 model outputs,
- 🔎 evidence,
- 🔥 explainability,
- 🧭 structured review,
- 👩⚕️ human interpretation.
The goal is not simply to expose model APIs.
The goal is to provide a coherent review experience.
The EndoXAI-RCT web prototype is accessible through Google Cloud Run:
https://endoxai-rct-459576379252.asia-south1.run.app/
| Component | Status |
|---|---|
| 🌐 Public web prototype | ✅ Available |
| ☁️ Google Cloud Run endpoint | ✅ Available |
| 🤗 Hugging Face Space | ✅ Available |
| 🧠 Research implementation | ✅ Repository source available |
| 👩⚕️ Autonomous clinical action | ❌ Not provided |
| 🏥 Medical-device status | ❌ Not claimed |
The associated Hugging Face Space is available at:
https://huggingface.co/spaces/janicecodes/EndoXAI-RCT
The Hugging Face environment represents part of the project's research and model-development history, while this GitHub repository is intended to provide the cleaner software/research repository for EndoXAI-RCT.
The repository currently contains implementation components associated with the following stack:
| Layer | Technology / Approach |
|---|---|
| 🌐 Web interface | HTML / CSS / JavaScript |
| ⚡ API | Python / FastAPI |
| 🧠 Deep learning | PyTorch-based model workflow |
| 🔥 Explainability | Grad-CAM-related implementation |
| 🖼️ Image processing | Python image-processing pipeline |
| 📦 Packaging | Docker |
| ☁️ Deployment | Google Cloud Run |
| 🤗 Research hosting | Hugging Face Spaces |
| 🔐 Configuration | Environment-based configuration |
The exact runtime configuration may evolve as the research prototype is refined.
A simplified view of the repository is:
EndoXAI-RCT/
│
├── assets/
│ ├── default-radiograph.jpg
│ └── endoxai-logo.png
│
├── models/
│
├── app.js
├── index.html
├── styles.css
│
├── fastapi_app.py
│
├── gradcam_resnet_lsl.py
├── make_clinical_reference_overlays.py
├── train_resnet_lsl.py
│
├── requirements.txt
├── Dockerfile
├── Run_Local_EndoXAI.bat
│
├── .env.example
├── .dockerignore
├── .gitignore
├── .gitattributes
│
└── README.md
Repository contents may evolve as deployment, model packaging and documentation are improved.
Large trained-model artifacts should not be committed directly to ordinary Git history.
For this reason, large .pt model files may be excluded through .gitignore.
For example:
*.pt
resnet_lsl_model.pt
This keeps the source repository lightweight and avoids GitHub file-size limitations.
The model artifact must therefore be obtained or provisioned separately when reproducing model-dependent functionality.
git clone https://github.com/Janicebenita/EndoXAI-RCT.git
cd EndoXAI-RCTpython -m venv .venv
.venv\Scripts\activatepython3 -m venv .venv
source .venv/bin/activatepip install -r requirements.txtUse the provided example configuration:
.env.example
Create a local .env where required.
Do not commit secrets, tokens or credentials.
The trained model artifact is intentionally not stored in ordinary Git history because of its size.
Place the required model in the path expected by the application before starting model-dependent inference.
Exact model-distribution instructions will be documented as repository packaging is finalized.
The repository includes:
Run_Local_EndoXAI.bat
for Windows-based local execution.
Alternatively, the FastAPI application can be started according to the application configuration.
A Dockerfile is included for containerized execution.
Typical workflow:
docker build -t endoxai-rct .followed by:
docker run -p 8080:8080 endoxai-rctPort and startup behaviour should follow the current
Dockerfileand application configuration.
EndoXAI-RCT separates:
SOURCE CODE
│
├── application logic
├── interface
├── inference pipeline
├── explainability logic
└── deployment configuration
from
LARGE MODEL ARTIFACTS
│
└── trained weights
This is intentional.
The source repository remains auditable and manageable while large artifacts can be versioned through an appropriate model-storage mechanism.
EndoXAI-RCT treats model evaluation and system evaluation as related but distinct concerns.
Examples include:
- predictive performance,
- class-specific behaviour,
- confidence,
- error characteristics,
- explainability outputs.
Examples include:
- whether the correct model is invoked,
- whether invalid input is rejected,
- whether evidence remains attributable,
- whether model roles are respected,
- whether failures are surfaced,
- whether explanations correspond to the appropriate output,
- whether the human-review boundary remains intact.
This distinction is important because:
EndoXAI-RCT should be interpreted as a research and engineering prototype.
It demonstrates architectural concepts including:
- multi-model orchestration,
- role-aware model usage,
- evidence-oriented output,
- explainability,
- failure awareness,
- software deployment,
- human-in-the-loop review.
It does not claim to demonstrate:
- autonomous diagnosis,
- autonomous treatment recommendation,
- regulatory approval,
- medical-device certification,
- unsupervised clinical deployment.
The EndoXAI-RCT architecture is associated with the research work:
EndoXAI-RCT: A Deployable Explainable AI Software Architecture for Multi-Model Clinical Image Review
The work focuses on the engineering architecture required to move beyond isolated AI predictions toward a structured multi-model review system.
Key architectural themes include:
- model-role management,
- primary-versus-advisory routing,
- evidence-aligned outputs,
- explainability,
- health-aware execution,
- failure-aware behaviour,
- artifact persistence,
- human-in-the-loop review.
International Conference on Recent Innovation in Science, Engineering and Technology — ICRISET 2026
Bibliographic information can be updated here once final publication metadata, proceedings information and DOI details are publicly available.
Many AI demonstrations stop at:
Dataset
↓
Training
↓
Accuracy
↓
Prediction
EndoXAI-RCT extends the engineering question:
Dataset / Models
↓
Model Responsibilities
↓
Software Integration
↓
Validation
↓
Multi-Model Execution
↓
Evidence Construction
↓
Routing
↓
Explainability
↓
Failure Handling
↓
Human Review
↓
Deployable Research Prototype
This is the architectural space that EndoXAI-RCT is designed to explore.
EndoXAI-RCT follows several important design principles.
|
AI output is presented for review rather than being treated as autonomous clinical action. |
The architecture aims to expose supporting model information rather than presenting unexplained conclusions. |
Invalid inputs and component failures should not silently become trusted evidence. |
|
Visual explanation supports inspection of model behaviour but does not prove a diagnosis. |
Model outputs should influence the workflow according to their intended architectural roles. |
The system is presented as an engineering/research prototype rather than a certified medical system. |
IMPORTANT
EndoXAI-RCT is a research and engineering prototype.
It is not a medical device.
It has not been presented here as a substitute for professional clinical judgement.
AI-generated predictions, confidence values, visualizations, heatmaps and other outputs must not be interpreted as independent medical advice or used as the sole basis for diagnosis or treatment.
Clinical interpretation must remain with appropriately qualified professionals operating within applicable clinical, ethical and regulatory requirements.
Public demonstrations of medical-image AI should avoid exposing identifiable patient information.
Users evaluating this repository should:
- use appropriately authorized or de-identified images,
- avoid committing patient images to Git,
- avoid placing credentials in source files,
- avoid exposing private datasets through public deployments,
- follow applicable institutional and regulatory requirements.
The repository is intended for research, software engineering and demonstration purposes.
For clarity, this repository does not claim:
❌ autonomous clinical diagnosis
❌ autonomous treatment selection
❌ replacement of dentists or specialists
❌ medical-device certification
❌ regulatory approval
❌ guaranteed clinical performance in uncontrolled environments
❌ that explainability heatmaps prove clinical causation
This distinction is intentional.
Potential future engineering directions include:
- stronger model packaging,
- additional model-role definitions,
- calibrated confidence handling,
- improved model-health monitoring,
- broader evaluation.
- additional explanation techniques,
- comparison of explanation methods,
- explanation-quality evaluation,
- reviewer-oriented visualization.
- stronger provenance tracking,
- persistent evidence objects,
- richer model disagreement representation,
- structured review histories.
- automated model artifact provisioning,
- stronger CI/CD,
- deployment health checks,
- versioned releases,
- improved observability.
- structured reviewer feedback,
- review audit trails,
- usability evaluation,
- expert-centered interface studies.
Although panoramic dental imaging is the application context used in EndoXAI-RCT, several architectural ideas are more general.
The pattern:
MULTIPLE AI MODELS
↓
DEFINED MODEL ROLES
↓
TRACEABLE EVIDENCE
↓
CONTROLLED ROUTING
↓
EXPLAINABILITY
↓
HUMAN REVIEW
can be investigated in other human-reviewed AI workflows where heterogeneous models contribute to a common decision-support process.
The broader contribution is therefore not limited to a single prediction task.
AI outputs should be accompanied by information that allows meaningful review.
A model's purpose should be understood before its output influences downstream logic.
Failure states should be identified explicitly rather than silently hidden.
Explainability should support scrutiny, not merely decorate a prediction.
The final interpretation remains outside autonomous model authority.
| Resource | Link |
|---|---|
| 🚀 Live EndoXAI-RCT | Launch Google Cloud Run |
| 🤗 Hugging Face Space | Open EndoXAI-RCT |
| 💻 GitHub Repository | Janicebenita/EndoXAI-RCT |
| 👩💻 Developer Profile | Janicebenita |
This repository may be useful to students, researchers and developers interested in:
- explainable AI,
- computer vision,
- clinical-AI software architecture,
- multi-model systems,
- evidence-grounded AI,
- human-in-the-loop AI,
- medical-image software engineering,
- model orchestration,
- AI deployment,
- responsible AI.
When reusing or extending the work, please preserve appropriate attribution and observe the repository's licensing terms.
If you use the architecture or research concepts from EndoXAI-RCT in academic work, please cite the associated publication once the final bibliographic record is available.
Temporary citation format:
@inproceedings{endoxairct2026,
title = {EndoXAI-RCT: A Deployable Explainable AI Software Architecture for Multi-Model Clinical Image Review},
author = {Janice Benita F. and co-authors},
booktitle = {International Conference on Recent Innovation in Science, Engineering and Technology},
year = {2026},
note = {ICRISET 2026}
}Replace the temporary entry with the official proceedings citation when final bibliographic metadata becomes available.
Research • Computer Vision • Explainable AI • Software Architecture • Human-in-the-Loop AI
Created as a research and engineering project by Janice Benita F
This software and associated documentation are provided for research, education and engineering demonstration.
Nothing in this repository should be interpreted as medical advice, clinical guidance, regulatory certification or authorization for clinical deployment.
Users are responsible for evaluating the suitability, security, privacy, regulatory and ethical requirements applicable to any derivative implementation.
🚀 Launch Live Platform
🤗 Explore on Hugging Face
💻 View GitHub Profile
Research Prototype · Explainable AI · Multi-Model Intelligence · Human-Governed Review