A focused workspace for AI-assisted image research. Analyse tissue patches and mammograms, then explore the model’s predictions, attention maps, and explanations.
Research & education. Professional review required.
Keep the source image in view. Inspect the evidence, then ask questions in context.
01
Upload a research image
Choose a modality and add a de-identified image. Select who the explanation is written for.
02
Review the output
Compare the source image with the model’s prediction, confidence score, and Grad-CAM attention map.
03
Ask a follow-up question
Use the conversation attached to the result to explore the explanation and its limitations.
TRANSPARENT BY DESIGN
The models. The data. The limits.
See the architecture, datasets, and project-reported evaluation results behind each module.
Histopathology
Test accuracy
88.0%
Histopathology
Test sensitivity
87.5%
Mammography · patient-level validation
Area under the ROC curve
0.8443
Project-reported results, not independently verified. Dataset performance does not establish clinical effectiveness.
YOUR NEXT STEP
Bring an image. Explore the output.
Research tools for tissue patches and mammograms.
MedAI / Workspace
Image analysis
Upload an image to inspect its model output.
Explanation for
Select an image to begin
Use de-identified research images. Images and any optional case context are sent to the configured backend. MedAI is not a diagnostic medical device.
Image workspaceH&E TISSUE
Choose an image or drop it here
PNG · JPG · TIFF · BMP
Grad-CAM overlay
Run analysis first
Mean activation
—
Peak activation
—
Raw heatmap
Greyscale activation map
▶Optional case context Sent with chat questions
No analysis yet
Upload a tissue patch, then select Analyse scan. The prediction and explanation will appear here.
AI prediction
Model confidence is not a diagnosis or a patient’s cancer risk.
Model confidence
Technical output / Raw logits
Logit · benign
—
Logit · malignant
—
Written explanation
template
Spatial analysis
AI assistant — ask follow-up questions about this scan
Clinician mode
Sample images / If installed
Tumor patch
Normal patch
Image workspaceMAMMOGRAPHY
Choose an image or drop it here
PNG · JPG · TIFF · DICOM (.dcm)
Grad-CAM overlay
Run analysis first
Raw heatmap
Your mammogram output will appear here.
Choose a mammogram, then select Analyse mammogram.
AI prediction
Model confidence is not a diagnosis or a patient’s cancer risk.
—
—
confidence
AUC 0.84 · Sens 70% · Spec 82% · RSNA validation
BI-RADS suggestion
Advanced — raw model outputs
Logit · benign
—
Logit · malignant
—
Written explanation
AI summary
AI assistant — ask about this mammogram
powered by gpt-oss-120b
Research & educational use only.
This EfficientNet-B4 ensemble (0.84 AUC, ~70% sensitivity / ~82% specificity on RSNA validation) is a research prototype — not a medical device and not a diagnosis. All outputs must be reviewed by a qualified radiologist.
Sample images / If installed
Cancer case
Benign case
MedAI / Conversation
Research assistant
Ask about the models, image analysis, or a result you have already generated.
MedAI / Research assistant
Models · Attention maps · Explanations
Quick topics
Image-specific questions
For questions about a particular image, use the conversation attached to its analysis result in the workspace.
MedAI / Model card
Models, data, and limitations.
A record of the architecture, training choices, and reported performance behind MedAI. These details are supplied by the project; they have not been independently verified in this frontend review.
Intended use: research and education. This prototype is not a diagnostic medical device. Dataset metrics do not establish performance in a clinical setting. Model predictions, attention maps, BI-RADS suggestions, and generated text require qualified professional review.
01 / HISTOPATHOLOGY
DenseNet-121
H&E tissue-patch classification. The project reports training on a deduplicated subset of PatchCamelyon.
Measure
Evaluation set
Reported value
Accuracy
Held-out test
88.0%
Sensitivity
Held-out test
87.5%
Sensitivity
Validation
90.3%
DenseNet-121 backbone
~7.2M-parameter CNN pretrained on ImageNet. Dense-block feature reuse across layers. Custom head: BN→Dropout(0.4)→Linear(1024→256)→ReLU→BN→Dropout(0.3)→Linear(256→2).
PatchCamelyon (PCam)
H&E patches from Camelyon16 slides. After MD5 deduplication: 220,025 unique samples (benign 130,908 / malignant 89,117 → pos_weight 1.469). Patient-disjoint train / val / test splits.
Breast-ROI cropping reduced AUC (~0.14 drop, aspect-ratio distortion); adding external CBIS-DDSM data hurt (film-vs-digital domain shift). The un-cropped, RSNA-only ensemble is final.
03 / EXPLANATIONS
Attention maps & text
Tools for inspecting model behaviour. Neither an attention map nor a generated explanation proves a finding.
Grad-CAM explainability
Gradients backpropagated w.r.t. the predicted class at the final conv block → global-average-pooled channel weights → weighted feature sum → ReLU → bilinear upsample to input resolution.
LLM explanation layer
FLAN-T5-large / BioMedLM / Llama 3.2 with a deterministic template engine. Three audience modes (clinician, researcher, patient) with genuinely different content.
What these outputs cannot establish
A model’s confidence score is not a patient’s probability of having cancer. Grad-CAM highlights model attention; it is not a validated tumour boundary. Generated explanations may contain errors. Suggested BI-RADS categories require radiologist review. Performance may change across hospitals, scanners, staining protocols, and patient populations.
Colour system. The RelixOrb logo inspires electric blue (#125BD3), cyan (#28D8FD), and violet (#473ADA). Deep navy (#0B1329) anchors typography, while blue-to-violet gradients highlight primary actions and cyan adds small accents. White and soft cool surfaces, consistent spacing, sans-serif typography, and clearly labelled states support the research workflow. Medical images and heatmaps retain their original colours.
Reference breast micrograph: “Pathologic mitosis” by M. Urezkova, CC BY 4.0. Converted to WebP and cropped by the layout. Reference only; not a MedAI inference or an endorsed use.
Design review: 8 October 2026. This is a frontend redesign. Original API configuration, requests, payloads, response handling, and model logic are retained. A complete accessibility audit and testing with intended users remain necessary.
REFERENCE IMAGE · NOT A MODEL OUTPUT
