MedLoop AI: A Local Human-in-the-Loop Explainable Medical Imaging System
MedLoop AI is a fully local Human-in-the-Loop medical imaging research system that explores how validated human feedback can be incorporated into successive AI model versions. It combines medical image classification, explainable AI, interactive annotation, human validation, iterative retraining, and model versioning within a single workflow.
MedLoop AI investigates a different approach to medical image annotation and model development: instead of treating human annotation as a one-time preprocessing step, the system treats validated human feedback as part of an iterative learning loop. The initial research focus is skin lesion analysis using the PAD-UFES-20 dataset. An AI model generates a prediction together with confidence and visual explanation. A human reviewer can then review, correct, and spatially annotate the case. Validated feedback is accumulated and used to construct subsequent training batches. Each newly trained model is maintained as a separate version and evaluated against a locked test set, allowing changes in performance to be measured across successive Human-in-the-Loop iterations. The entire pipeline is designed to run locally, including image processing, inference, explainability, annotation, training, evaluation, model versioning, and data storage.
Method
The core workflow is: Upload → Assign → Train → Predict → Explain → Human Review → Correct → Validate → Accumulate → Retrain → Evaluate → Version → Deploy → Repeat The initial classification model uses EfficientNet-B0 with PyTorch. Grad-CAM is used to provide visual explanations and support coarse localization of relevant image regions. During human review, reviewers can validate or correct the predicted disease label and provide spatial annotations using bounding boxes, polygons, or round-box annotations. Validated samples are accumulated into Human-in-the-Loop training batches. When the configured retraining threshold is reached, a new candidate model can be trained locally. Each candidate model is evaluated against the same locked test set and compared with the currently active model. This enables systematic measurement of whether successive cycles of validated human feedback lead to measurable changes in model performance. The system also maintains model versions, training batches, AI predictions, human corrections, annotations, and evaluation metrics to support traceability and reproducible experimentation.
MedLoop AI: A Local Human-in-the-Loop Explainable Medical Imaging System for Interactive Annotation and Continuous Model Refinement