Histopathology · Mammography · Explainable AI

AI-powered breast cancer intelligence platform

Research-grade breast cancer detection across two modalities — histopathology and mammography — with explainable AI. Upload a tissue patch or a mammogram and receive a prediction, a spatial heatmap, and a natural-language report in seconds.

Histopathology module

88.0%

Test accuracy

87.5%

Sensitivity

Mammography module

0.8443

Patient-level AUC

SCAN VIEWER

Histopathology

Histopathology · H&E patch

Mammography

Mammography · DICOM view

SAMPLE RESULT — ILLUSTRATIVE

PREDICTION

⚠ MALIGNANT

CONFIDENCE

0.94

Platform capabilities

Two detection modules, one explainable platform

Histopathology module

DenseNet-121

~7.2M-parameter CNN fine-tuned on 220K deduplicated PCam H&E tissue patches.

StainJitter (H&E)

Stain augmentation in HED colour space, reducing lab-specific staining overfitting.

Sensitivity-first

Checkpoints selected by validation sensitivity — tuned to catch malignant patches.

Mammography module

EfficientNet-B4 ensemble

Three EfficientNet-B4 models (seeds 42 / 123 / 999) trained on RSNA data, averaged at inference.

DICOM + VOI LUT

Native DICOM loading with VOI-LUT windowing. Also accepts PNG, JPG and TIFF.

BI-RADS scoring

Suggested BI-RADS category (2–5) from malignancy probability. Radiologist confirmation required.

Shared platform layers — both modules

Grad-CAM heatmaps

Spatial attention maps showing which regions drove each prediction.

LLM explanation

FLAN-T5 / BioMedLM / Llama 3.2 reports in clinician, researcher or patient modes.

Chat assistant

Ask follow-up questions about any result, in clinician, researcher or patient mode.

Confidence scores

Calibrated probability scores with raw logits and decision margin.

How it works

From scan to report in five steps

1

Upload a scan

Histology patch (PNG / JPG / TIFF) or mammogram (DICOM / PNG) — drag and drop onto the dashboard

2

AI detection

Routed by modality — DenseNet-121 for histology patches, the EfficientNet-B4 ensemble for mammograms

3

Grad-CAM overlay

Spatial heatmap generated showing the regions that drove the prediction

4

LLM explanation

Audience-tailored natural language report

5

Chat and review

Ask follow-up questions in the inline chat assistant

Analysis dashboard

Upload a histopathology patch to run the full pipeline

Audience:

▶ Patient record (optional — improves chat responses)

📤

Drop a histopathology patch here

PNG · JPG · TIFF · BMP

🤖

Upload a scan and click Analyse

Results appear here

Try a sample — click to analyse

Tumor patch

Normal patch

Drop a mammogram here

PNG · JPG · TIFF · DICOM (.dcm)

Upload a mammogram and click Analyse

Supports DICOM, PNG, and JPG

EfficientNet-B4

19M parameter CNN scaled for 512×512 full-field mammography. RSNA 2022 competition-proven architecture.

BI-RADS scoring

Automatic BI-RADS category suggestion (2–5) based on malignancy probability. Requires radiologist confirmation.

DICOM support

Native DICOM (.dcm) loading with VOI LUT windowing. Also accepts PNG, JPG, and TIFF formats.

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.

Try a sample — click to analyse

Cancer case

Benign case

🤖

AI radiology assistant

DenseNet-121 · FLAN-T5 · Grad-CAM

Quick topics

› DenseNet-121 architecture

› StainJitter augmentation

› OneCycleLR scheduler

› PCam deduplication

› Mixup augmentation

› Training improvements

Connect live scan

Run an analysis in the dashboard first, then return here to ask questions about that specific scan.

Model card

Architecture, datasets & methodology

Histopathology module

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.

Training methodology

OneCycleLR (max_lr 3e-3, pct_start 0.3) · Mixup (α 0.4) · StainJitter (HED, strength 0.05) · label smoothing 0.1 · checkpoint by validation sensitivity.

Histopathology results

Held-out test: 88.0% accuracy, 87.5% sensitivity (val sensitivity 90.3%, best epoch 13/20). Improvement vs baseline +6.8pp (80.7% → 87.5%).

Mammography module

EfficientNet-B4 ensemble

Three EfficientNet-B4 models (seeds 42 / 123 / 999), probabilities averaged at inference. Member AUCs 0.7989 / 0.8254 / 0.8083.

RSNA dataset & preprocessing

RSNA Screening Mammography (2022). DICOM decoding with VOI-LUT windowing, resized for full-field input. Held-out patient-level validation split.

Mammography results

Ensemble patient-level AUC 0.8443 · sensitivity 70.1% · specificity 82.4% (threshold 0.50). Suggested BI-RADS 2–5 derived from malignancy probability.

Negative ablations

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.

Shared layers — both modules

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.