ThyroLens AI

ThyroLens AI is a research and educational AI decision-support prototype for binary thyroid image classification. It combines classification, explainability, calibration, image-quality checks, and feature-space similarity into one analysis workflow.

Medical imaging AI outputs can be difficult to interpret and may not communicate whether an input is technically suitable or sufficiently similar to the data used during development.

Combining prediction with explainability and technical reliability signals encourages more cautious, transparent, and human-reviewed use of AI outputs.

Designed and implemented the classification application, explainability workflow, image-quality checks, input-similarity screening, calibration integration, and technical reliability assessment.

JPEG/PNG upload -> RGB preprocessing -> FibonacciNet prediction -> Platt calibration and feature-space similarity -> image-quality assessment -> technical reliability -> Grad-CAM visualization -> web response or DOCX report.

Python, TensorFlow, FibonacciNet, FastAPI, Grad-CAM, Platt scaling, Mahalanobis distance, Streamlit, DOCX reporting

Built a modular analysis pipeline that separates raw model output, calibrated probability, image-quality signals, input similarity, and technical reliability. Added Grad-CAM visualization, report generation, and reusable evaluation utilities.

• Kept raw model score, calibrated probability, and technical reliability as separate quantities.
• Used deterministic image-quality heuristics to identify technically weak inputs.
• Added feature-space Mahalanobis distance for input-similarity and OOD screening.
• Made reliability weights and thresholds configurable.
• Preserved explicit safety boundaries instead of presenting the prototype as a clinical diagnostic system.

• Communicating model confidence without overstating medical certainty.
• Combining heterogeneous technical signals into an interpretable reliability assessment.
• Handling image-quality and out-of-distribution concerns alongside classification.
• Documenting limitations caused by unavailable historical split and dataset provenance details.

• Separated raw score, calibrated probability, and AI technical reliability in the interface.
• Added quality, similarity, and calibration components to the reliability pipeline.
• Used Grad-CAM as an approximate attention visualization with clear limitations.
• Included model-card documentation and human-in-the-loop recommendations.

Delivers an end-to-end research prototype that produces thyroid image classification results alongside explainability, image-quality signals, input-similarity screening, calibration information, and an engineering-level technical reliability assessment.

The repository includes reusable evaluation utilities for held-out predictions. No independently verified clinical performance metrics are claimed because the historical split manifest and independent validation are unavailable.

• A strong classifier score alone is not sufficient to establish technical reliability.
• Calibration and input-similarity signals add important context to model outputs.
• Explainability visualizations should be communicated as approximate model-attention aids, not causal or clinical evidence.
• Clear model limitations are essential when working with medical imaging prototypes.

• Persist a versioned, deduplicated dataset and independent test split.
• Perform external and prospective validation with appropriate governance.
• Evaluate calibration and OOD behavior across acquisition devices and sites.
• Add a research dashboard only when real evaluation artifacts are available.