Kidney Stone Images Classification

Kidney Stone Images Classification/Dataset/README.md

@@ -0,0 +1,55 @@
+# Dataset — Kidney Stone Images
+
+## Source
+**Kaggle:** https://www.kaggle.com/datasets/safurahajiheidari/kidney-stone-images
+
+## About the Dataset
+- **Total Images:** ~1,299 CT scan images (Roboflow Universe export)
+- **Classes:** `nc: 1` — single class `Tas_Var` (Turkish: "Stone Present") — all images contain stones
+- **Task:** Binary classification — **Small Stone vs Large Stone** (median bounding box area split, threshold = 0.0015)
+- **Annotation Format:** YOLO format (bounding box labels per image)
+- **Splits:** Pre-divided into `train`, `valid`, and `test` (1054 / 123 / 123)
+
+## Directory Structure (after download)
+```
+/kaggle/input/datasets/safurahajiheidari/
+└── kidney-stone-images/
+    ├── train/
+    │   ├── images/   ← .jpg CT scan images
+    │   └── labels/   ← .txt YOLO annotation files
+    ├── valid/
+    │   ├── images/
+    │   └── labels/
+    ├── test/
+    │   ├── images/
+    │   └── labels/
+    ├── data.yaml
+    ├── README.dataset.txt
+    └── README.roboflow.txt
+```
+
+## Label Format (YOLO)
+Each `.txt` label file contains one line per annotated stone:
+```
+class_id  x_center  y_center  width  height
+```
+All values (except `class_id`) are normalised to [0, 1].  
+`class_id` is always `0` (`Tas_Var`). Multiple lines = multiple stones in the image.
+
+## Classification Strategy
+Since all images contain stones (`nc: 1`), classification uses the **largest bounding box area** per image:
+
+- `max_area >= median (0.0015)` → **`large`** stone (label = 1)
+- `max_area < median (0.0015)` → **`small`** stone (label = 0)
+
+This guarantees a perfectly balanced 50/50 split and provides visually distinct classes — large stones occupy more of the CT image.
+
+## How to Add in Kaggle
+1. Open your Kaggle Notebook
+2. Click **+ Add Data** (top right)
+3. Search: `kidney-stone-images safurahajiheidari`
+4. Click **Add**
+5. Dataset mounts at `/kaggle/input/datasets/safurahajiheidari/kidney-stone-images/`
+
+> **Note:** Set `DATASET_PATH = '/kaggle/input/datasets/safurahajiheidari/kidney-stone-images'` in Cell 1.
+```

Kidney Stone Images Classification/Model/README.md

@@ -0,0 +1,89 @@
+# Kidney Stone Images Classification using Deep Learning
+
+## 🎯 Aim of the Project
+To classify CT scan images from the kidney stone dataset using deep learning and compare four transfer learning models by performance metrics to identify the best approach for stone size classification.
+
+## 📂 Dataset
+- **Source:** [Kaggle — Kidney Stone Images with Bounding Box Annotations](https://www.kaggle.com/datasets/safurahajiheidari/kidney-stone-images)
+- **Total Images:** ~1,299 CT scan images
+- **Splits:** Pre-divided into `train / valid / test` (1054 / 123 / 123)
+- **Label format:** YOLO `.txt` annotations (bounding boxes), `nc: 1`, class `Tas_Var`
+- **Classification task:** Binary — **Small Stone vs Large Stone** (median bounding box area split, threshold = 0.0015)
+
+> All images contain kidney stones (`nc: 1`). Classification is based on the largest bounding box area per image — images above the dataset median are labelled `large`, below are labelled `small`. Large stones visually dominate more of the CT image, providing more learnable signal than stone count.
+
+## 🧠 Models Used
+
+| Model | Key Strength |
+|---|---|
+| **Xception** | Depthwise separable convolutions — captures subtle stone texture patterns |
+| **InceptionV3** | Multi-scale inception modules — handles variation in stone size across the image |
+| **EfficientNetV2S** | Fused-MBConv + progressive learning — outperforms V1 on small medical datasets |
+| **ConvNeXt-Tiny** | Modern (2022) pure-CNN — matches transformer accuracy with standard convolutions |
+
+All models use **two-phase transfer learning**: frozen base (phase 1) → fine-tuning top 30 layers at `lr=1e-5` (phase 2).
+
+## 🔬 Approach
+
+1. **Dataset Structure Check** — Auto-detect path, walk directory, read `data.yaml`
+2. **YOLO Label Parsing** — Extract bounding boxes, compute max area per image
+3. **EDA** — Class distribution, bounding box visualisation, stone size distribution, augmentation preview
+4. **Median Split Labelling** — `large` if `max_area ≥ 0.0015`, else `small` — guarantees balanced classes
+5. **tf.data Pipeline** — AUTOTUNE, caching, shuffling, augmentation (flip, brightness, contrast)
+6. **Class Weights** — Applied during training as best practice
+7. **Training** — Each model trained independently; histories saved with `pickle` after every model
+8. **Evaluation** — Test set evaluation using Accuracy, Precision, Recall, F1-Score, ROC-AUC, Cohen's Kappa
+9. **Visualisation** — Training curves, confusion matrices, ROC curves, multi-metric bar chart
+
+## 📊 Exploratory Data Analysis
+
+![EDA Overview](../Images/eda_overview.png)
+
+![Bounding Boxes](../Images/eda_bboxes.png)
+
+![Class Imbalance](../Images/eda_imbalance.png)
+
+![Augmented Samples](../Images/eda_augmented_samples.png)
+
+## 📈 Training Curves
+
+![Training Curves](../Images/training_curves.png)
+
+## 🔲 Confusion Matrices
+
+![Confusion Matrices](../Images/confusion_matrices.png)
+
+## 📉 ROC Curves
+
+![ROC Curves](../Images/roc_curves.png)
+
+## 🏆 Performance Comparison
+
+![Final Comparison](../Images/final_comparison.png)
+
+| Model | Accuracy | Precision | Recall | F1-Score | ROC-AUC | Cohen κ |
+|---|---|---|---|---|---|---|
+| **ConvNeXtTiny** | **0.7154** | **0.7201** | **0.7148** | **0.7136** | **0.7501** | **0.4302** |
+| InceptionV3 | 0.6911 | 0.6918 | 0.6908 | 0.6905 | 0.7496 | 0.3817 |
+| EfficientNetV2S | 0.5935 | 0.5937 | 0.5936 | 0.5935 | 0.6214 | 0.1872 |
+| Xception | 0.5447 | 0.5447 | 0.5444 | 0.5440 | 0.5933 | 0.0889 |
+
+## ✅ Conclusion
+
+**Best Model: ConvNeXtTiny** — Accuracy: 71.54% | F1-Score: 0.7136 | ROC-AUC: 0.7501 | Cohen κ: 0.4302
+
+ConvNeXtTiny and InceptionV3 are closely matched on ROC-AUC (0.7501 vs 0.7496), but ConvNeXtTiny edges ahead on all other metrics. Its modern large-kernel convolution design likely captures the spatial spread of larger stones more effectively.
+
+EfficientNetV2S and Xception underperformed, suggesting that depthwise and compound-scaled architectures need more data to generalise well on this subtle medical imaging task.
+
+All models were evaluated beyond plain accuracy using ROC-AUC and Cohen's Kappa — metrics designed for balanced binary tasks in medical imaging.
+
+## 📦 Libraries Used
+- TensorFlow / Keras
+- NumPy, Pandas
+- Matplotlib, Seaborn
+- scikit-learn (classification_report, roc_auc_score, cohen_kappa_score)
+- pickle (history persistence)
+
+## 🖥️ Platform
+Kaggle Notebook — T4 GPU (2×)