Himel000/kidney-tumor-efficientnetb4

Kidney Tumor Identification — EfficientNetB4

Model Description

This model classifies kidney CT scan images into four categories using EfficientNetB4 with transfer learning and a custom classification head. It is part of the Kidney Tumor Identification System — a production-ready, explainable AI application for medical image analysis.

⚠️ Medical Disclaimer: This model is intended for research and educational purposes only. It is NOT a clinical diagnostic tool. Always consult a qualified radiologist or physician for medical decisions.

---

Model Details

Property	Value
Architecture	EfficientNetB4 + Custom Classification Head
Input Size	380 × 380 × 3
Output Classes	4 (Cyst, Normal, Stone, Tumor)
Pretrained Weights	ImageNet
Framework	TensorFlow / Keras
Training Strategy	Two-phase transfer learning
Preprocessing	tf.keras.applications.efficientnet.preprocess_input

Model Architecture

Input (380, 380, 3)
    ↓
EfficientNetB4 (frozen in phase 1, last 4 layers unfrozen in phase 2)
    ↓
GlobalAveragePooling2D
    ↓
BatchNormalization
    ↓
Dropout (0.3)
    ↓
Dense (4, softmax) → [Cyst, Normal, Stone, Tumor]

---

Training Details

Dataset

• Name: CT KIDNEY DATASET — Normal, Cyst, Tumor and Stone
• Source: Kaggle
• Total Images: 12,446
• Split: 70% train / 15% validation / 15% test

Class	Count
Normal	5,077
Cyst	3,709
Tumor	2,283
Stone	1,377

Training Configuration

Parameter	Phase 1	Phase 2
Epochs	10	20
Learning Rate	0.001	0.0001
Batch Size	32	32
Base Model	Frozen	Last 4 layers unfrozen
Optimizer	Adam	Adam
Loss	Categorical Crossentropy	Categorical Crossentropy

Data Augmentation (Training only)

• Random horizontal flip
• Random rotation (±15°)
• Random zoom (±10%)
• Random brightness adjustment (±15%)

Hardware

• GPU: Tesla T4 × 2 (Kaggle Notebooks)
• Training time: ~1.5 hours (phase 1 + phase 2)

---

Evaluation Results

Test Set Performance

Metric	Value	Threshold
Validation Accuracy	99.46%	—
Test Accuracy	85.36%	—
Test AUC-ROC	98.38%	≥ 90% ✅
Test Sensitivity	85.36%	≥ 85% ✅
Test Specificity	95.00%	—
Test F1 Score	85.25%	—

Evaluation Gate: Model is only pushed to production if AUC-ROC ≥ 0.90 AND Sensitivity ≥ 0.85. This gate is enforced automatically in the CI/CD pipeline.

MLflow Experiment Tracking

All training runs are logged on DagsHub: 🔗 https://dagshub.com/himelds/kidney-tumor-identification-system.mlflow

---

How to Use

Quick inference

from huggingface_hub import hf_hub_download
import tensorflow as tf
import numpy as np
from PIL import Image

# Load model
model_path = hf_hub_download(
    repo_id="Himel000/kidney-tumor-efficientnetb4",
    filename="model.keras"
)
model = tf.keras.models.load_model(model_path)

# Class names (model output order)
CLASS_NAMES = ["Cyst", "Normal", "Stone", "Tumor"]

# Preprocess image
def preprocess(image_path):
    img = Image.open(image_path).convert("RGB").resize((380, 380))
    arr = np.array(img, dtype=np.float32)
    arr = tf.keras.applications.efficientnet.preprocess_input(arr)
    return np.expand_dims(arr, axis=0)

# Predict
image = preprocess("ct_scan.jpg")
predictions = model.predict(image)
predicted_class = CLASS_NAMES[np.argmax(predictions)]
confidence = float(np.max(predictions))

print(f"Prediction: {predicted_class} ({confidence:.2%} confidence)")

Via REST API

import requests

response = requests.post(
    "https://himel000-kidney-tumor-api.hf.space/api/v1/predict",
    files={"file": open("ct_scan.jpg", "rb")}
)

result = response.json()
print(f"Class: {result['predicted_class']}")
print(f"Confidence: {result['confidence']:.2%}")
print(f"Uncertain: {result['is_uncertain']}")
# result also contains: gradcam_base64, probabilities, uncertainty_score

Via Live Demo

👉 https://kidney-tumor-identification-system.streamlit.app/

---

Explainability

This model is deployed with full explainability support:

• Grad-CAM — highlights regions of the CT scan the model focused on
• Monte Carlo Dropout — runs 20 inference passes to estimate prediction uncertainty
• Uncertainty flag — predictions with uncertainty score > 0.3 are flagged for radiologist review

---

Limitations

• Trained on a single public dataset — may not generalize to CT scans from all machines or hospitals
• Not validated in a clinical setting
• Not FDA/CE cleared
• Performance may vary on images with different contrast, resolution, or windowing settings
• Class imbalance in training data (Normal: 5,077 vs Stone: 1,377) may affect performance on underrepresented classes

---

Intended Use

Appropriate uses

• Medical AI research
• Computer vision education
• Benchmarking kidney CT classification models
• Prototype for hospital-specific retraining

Inappropriate uses

• Standalone clinical diagnosis
• Replacement for radiologist evaluation
• Any use without appropriate medical supervision

---

Citation

If you use this model in your research, please cite:

@software{das2026kidney,
  author = {Das, Himel},
  title = {Kidney Tumor Identification System},
  year = {2026},
  url = {https://github.com/himelds/kidney-tumor-identification-system},
  note = {EfficientNetB4-based CT scan classifier with Grad-CAM explainability}
}

Dataset citation:

@dataset{islam2022kidney,
  author = {Islam, Nazmul},
  title = {CT KIDNEY DATASET: Normal-Cyst-Tumor and Stone},
  year = {2022},
  publisher = {Kaggle},
  url = {https://www.kaggle.com/datasets/nazmul0087/ct-kidney-dataset-normal-cyst-tumor-and-stone}
}

---

Author

Himel Das

---

License

MIT License — see LICENSE for details.