ablackwellb/efficientnetb0-ct

Attribution

Original Source:

Hany H. (2020). Chest CT-Scan Images Dataset. Kaggle.
https://www.kaggle.com/datasets/hanyhossam/chest-ctscan-images-dataset

Original License:

Database: Open Database Commons Open Database License (ODbL v1.0)
https://opendatacommons.org/licenses/odbl/1-0/

Derived Dataset Author:

Ashley Blackwell (2025). Chest CT-Scan Images (Cleaned, Derived from Hany et al.). Hugging Face Datasets.
https://huggingface.co/datasets/ashleyblackwell/lung-ct-cleaned-hany

---

Cleaning & Preprocessing Summary

The original dataset was processed and curated to ensure consistency, quality, and reproducibility for use in deep-learning experiments (i.e.., the EfficientNet-B0 Lung CT Classifier).

Steps Performed

1. Integrity Checks: Removed corrupted or unreadable .jpg and .png files.
2. Resolution Standardization: Resized all images to 224 × 224 × 3 pixels.
3. Color Normalization: Converted grayscale scans to RGB format.
4. Class Organization: Verified folder structure for four diagnostic categories:
   • Adenocarcinoma
   • Large-Cell Carcinoma
   • Squamous-Cell Carcinoma
   • Normal
5. Stratified Splits:
   • Train: 70%
   • Validation: 20%
   • Test: 10%
6. Metadata File: Generated metadata.csv containing filename, class label, and original resolution for traceability.

---

Dataset Overview

Split	%	Notes
Train	70%	Stratified by class
Validation	20%	For hyperparameter tuning
Test	10%	Final evaluation set
Total	100%	All cleaned and standardized

---

Intended Use

• Purpose:
  Designed for research, coursework, and educational demonstrations in medical image classification, model interpretability (Grad-CAM), and reproducible machine learning pipelines.

• Out of Scope:
  This dataset must not be used for clinical diagnosis, treatment decisions, or commercial medical software development.

---

Legal & License Information

License

This dataset is distributed under the Open Data Commons Open Database License (ODbL v1.0).
You are free to:

• Share: Copy, distribute, and use the database.
• Create: Produce works from the database.
• Adapt: Modify, transform, and build upon the database.

Full legal text:
https://opendatacommons.org/licenses/odbl/1-0/

---

Intended Use

• Purpose:
  Designed for research, coursework, and educational demonstrations in medical image classification, model interpretability (Grad-CAM), and reproducible machine learning pipelines.

Scope

• Intended: Research, UMGC coursework, model-interpretability demos (Grad-CAM), benchmarking.

Out-of-scope: Clinical diagnosis, patient triage, or any safety-critical application.

• Model Architecture
• Backbone: EfficientNet-B0 (ImageNet-initialized, fine-tuned)
• Input size: 224 × 224 × 3
• Head: GlobalAveragePooling → Dropout (TODO: rate) → Dense(4, softmax)
• Loss: Categorical Cross-Entropy
• Optimizer: TODO (e.g., Adam, lr = 1e-4 with decay)
• Epochs / Batch size: TODO
• Class labels (index): 0: Adenocarcinoma 1: Large-Cell Carcinoma 2: Squamous-Cell Carcinoma 3: Normal

---

Data & Preprocessing

Source: Derived from Hany Lung Cancer CT Scan dataset (Kaggle). Corrupted and irregular-resolution images were removed and all remaining images standardized to 224×224. Split: Train/Val/Test = 70/20/10 (stratified). Transforms: Resize → RGB conversion → normalize to [0,1] or use preprocess_input. Artifacts logged: Confusion matrix, classification report, Grad-CAM overlays. Attribution: Credit original dataset per its license when sharing or publishing.

---

Evaluation

Test set size: TODO:N Metrics (macro): Accuracy, Precision, Recall, F1 Class Precision Recall F1 Support Adenocarcinoma TODO TODO TODO TODO Large-Cell TODO TODO TODO TODO Squamous TODO TODO TODO TODO Normal TODO TODO TODO TODO Macro Avg TODO TODO TODO N

Suggested Environment

tensorflow==2.15.0 keras==2.15.0 huggingface_hub>=0.23.0 numpy>=1.24

---

Explainability (Grad-CAM)

Last conv layer: top_conv for EfficientNet-B0. Tip: Use Grad-CAM to overlay heatmaps and validate that the model focuses on pathologically relevant regions.

Limitations, Bias & Ethical Considerations

Domain shift: CT protocols and scanners vary; may affect generalization.

Label noise: Community datasets can contain mislabels. Generalization: Model is not clinically validated. Mitigation: Use Grad-CAM audits and external validation before any applied use.

---

Training & Reproducibility

Hardware: Local GPU Seed / Determinism: 42 Reproduction steps: (https://github.com/ablackwellb/cancer-site-detection-cnn/blob/main/cancer_site_detection_cnn.ipynb)

Training Data

This model relies on the original source data collected and curated by Hany Hossam:

• Original Source: Hany H. (2020). Chest CT-Scan Images Dataset. Kaggle. Dataset Link
• Dataset License: Open Data Commons Open Database License (ODbL v1.0)

Dataset BibTeX Citation

@dataset{hany2020chestct,
  author    = {Hany, H.},
  title     = {Chest CT-Scan Images Dataset},
  year      = {2020},
  publisher = {Kaggle},
  url       = {https://kaggle.com}
}

License

Original Dataset: odbl-1.0 Dataset (derived): Follow the original dataset’s license terms and provide credit to the creator.

Citation

If you use this dataset, please cite both the original source and the derived version:

Original dataset:

Hany H. (2020). Chest CT-Scan Images Dataset. Kaggle.
https://www.kaggle.com/datasets/hanyhossam/chest-ctscan-images-dataset

Derived version:

Blackwell, A. (2025). Chest CT-Scan Images (Cleaned, Derived from Hany et al.) [Dataset]. Hugging Face.
https://huggingface.co/datasets/ashleyblackwell/lung-ct-cleaned-hany

---

How to Use (Load & Inference)

Option A — Download from the Hub

• from huggingface_hub import hf_hub_download import json, numpy as np, tensorflow as tf from tensorflow.keras.preprocessing import image

REPO_ID = "TODO:your-username/efficientnetb0-lung-ct-4class"

model_path = hf_hub_download(repo_id=REPO_ID, filename="model.keras") class_map_path = hf_hub_download(repo_id=REPO_ID, filename="class_map.json")

model = tf.keras.models.load_model(model_path, compile=False) with open(class_map_path) as f: idx_to_label = json.load(f)

def preprocess(img_path): img = image.load_img(img_path, target_size=(224, 224)) x = image.img_to_array(img) x = np.expand_dims(x, 0) x = x / 255.0 # or use tf.keras.applications.efficientnet.preprocess_input(x) return x

x = preprocess("path/to/ct_slice.png") probs = model.predict(x, verbose=0)[0] for i, p in enumerate(probs): print(f"{idx_to_label[str(i)]}: {p:.3f}") print("Predicted:", idx_to_label[str(int(np.argmax(probs)))])

Option B — Snapshot Download (Local Folder)

• from huggingface_hub import snapshot_download local_dir = snapshot_download(repo_id="TODO:your-username/efficientnetb0-lung-ct-4class")

loads ./model.keras and ./class_map.json from local_dir

---