The project presents a hybrid augmentation framework combining SMOTE (Synthetic Minority Oversampling Technique) and ACGAN (Auxiliary Classifier Generative Adversarial Network) to balance CT scan datasets for chronic kidney disease (CKD) detection.
Used transfer learning models (VGG16, MobileNetV2, EfficientNetV2) to classify kidney images as Normal, Cyst, Tumor, or Stone.
Used the publicly available CT KIDNEY DATASET: Normal-Cyst-Tumor and Stone from Kaggle:
https://www.kaggle.com/datasets/nazmul0087/ct-kidney-dataset-normal-cyst-tumor-and-stone
| Category | Sample Image |
|---|---|
| Normal Kidney |  |
| Cyst |  |
| Tumor |  |
| Stone |  |
Our approach combines SMOTE and ACGAN for hybrid data augmentation.
-
SMOTE (Synthetic Minority Oversampling Technique)
- Generates new feature-space samples for underrepresented classes.
- Improves class balance without duplicating data.
-
ACGAN (Auxiliary Classifier GAN)
- Generates label-conditioned CT images with improved anatomical fidelity.
- Enhances model generalization across kidney states.
-
Classification Models
- VGG16, EfficientNetV2, and MobileNetV2 were fine-tuned.
- Compared across original and augmented datasets.
| Model | Original Dataset Accuracy | Synthesized Dataset Accuracy | Training Duration |
|---|---|---|---|
| VGG16 | 99.2% | 97.3% | ~150 min |
| EfficientNetV2 | 97.6% | 97.0% | ~455 min |
| MobileNetV2 | 97.2% | 95.1% | ~111 min |
Hybrid augmentation improved model robustness and reduced overfitting, producing clinically relevant accuracy scores.