AidRenal · CT Kidney Scan

CT Kidney Classifier

Upload a renal CT scan · Select architecture · Classify into Normal / Cyst / Stone / Tumor

Source: scores.json · auto-refreshed from /scores · params.yaml MODEL_NAME: EfficientNetV2B3

Input · Renal CT Scan224 × 224

⬡

Drop renal CT scan or click to upload

Greyscale abdominal/renal CT only · JPEG or PNG

Output · Classification DenseNet121

◎

Upload a renal CT scan to classify

Renal finding

—

Confidence: —

⚠ Research instrument only · 4-class renal CT classifier · Consult a nephrologist or urologist

Model Performance

Validated results from DagsHub MLflow · all 4 architectures complete

Best Val Accuracy

99.96%

DenseNet121 & ResNet50V2

scores.json Accuracy

—

Live push · EfficientNetV2B3

Dataset

12,446

Cyst · Normal · Stone · Tumor

MLflow Runs

18+

Tracked on DagsHub

Model Cards

DenseNet121

7M params · Adam · GAP head · 20 layers unfrozen

99.96%

● complete · epoch 42 · loss 0.0044

ResNet50V2

25M params · Adam · GAP head · 20 layers unfrozen

99.96%

● complete · epoch 43 · loss 0.0058

EfficientNetV2B3

14M params · Adam · GAP head · 20 layers unfrozen

96.08%

● complete · epoch 7 · loss 0.1180

VGG16

138M params · SGD · Flatten head · 0 unfrozen

65.94%

● complete · epoch 150 · loss 0.969

Live Per-Class Scores (scores.json)

scores.json · current model push auto-refresh

Loading live scores…

Head-to-Head Comparison

Actual MLflow metrics from DagsHub · renal CT 4-class classification

Metric	DenseNet121	ResNet50V2	EfficientNetV2B3	VGG16
Val Accuracy	99.96%	99.96%	96.08%	65.94%
Val Loss	0.0044	0.0058	0.1180	0.9692
Macro F1	~1.000	~1.000	~0.960	0.659
Stone Recall	~100%	~100%	~100%	~66%
Tumor Recall	~100%	~100%	~100%	~69%
Parameters	7M	25M	14M	138M
Best Epoch	42	43	7	150+
Optimizer	Adam	Adam	Adam	SGD
Pooling Head	GAP	GAP	GAP	Flatten
Unfrozen Layers	20	20	20	0
LR	0.0001	0.0001	0.0001	0.0003
Batch Size	32	32	32	32

Accuracy vs Parameters (M)

Experiment Log

All MLflow runs · renal CT classification · tracked on DagsHub

Full tracking: dagshub.com/MicroPyscho/renal_tumour_project.mlflow ↗

Confusion Matrix

VGG16 best run · renal CT val set · hover cells for anatomy detail

        Actual → rows  |  Predicted → cols

Per-class Recall

Clinical Risk

Cost-weighted misclassification · renal oncology priority framework

Risk Cost Matrix (0=correct, 10=worst)

Danger Pairs · Renal Context

⚠ RENAL ONCOLOGY NOTE — Predicting renal Tumour as Normal is the highest-risk misclassification — a missed RCC (renal cell carcinoma) can progress from resectable to metastatic. DenseNet121 and ResNet50V2 at 99.96% accuracy with ~100% Tumour recall dramatically reduce this risk. Target ≥99% Tumour recall and ≥98% sensitivity before any clinical consideration.

Grad-CAM · Renal Anatomy

Simulated activation maps on kidney anatomy regions · replace with real tf-explain output

Attention Heatmap · Kidney Cross-section Tumour

Kidney Region Activation

Real implementation:


        from tf_explain.core import GradCAM; GradCAM().explain((val_data, None), model, class_index=CLASS_NAMES.index("Tumor"))

Classes: 0=Cyst 1=Normal 2=Stone 3=Tumor · Layer: block5_conv3 (VGG16) / last conv block (DenseNet/ResNet/EfficientNet)