Lung Nodule Segmentation and Detection

---

Project Overview

This project focuses on the automated detection of pulmonary nodules in CT scans using the LUNA16 dataset. We employ:

• U-Net for precise segmentation of lung nodules.
• Convolutional Neural Networks (CNNs) for classification and false positive reduction.

The goal is to build an end-to-end pipeline that accurately segments nodules and classifies candidates to aid in early lung cancer detection.

---

Dataset

We use the publicly available LUNA16 (LUng Nodule Analysis 2016) dataset:

• LUNA16 Part 1 – Zenodo
• LUNA16 Part 2 – Zenodo

Please download both parts and extract them before running the preprocessing and training notebooks.

---

How to Proceed

• You can use VS Code for local development and training.
• For faster training, Google Colab with TPU is used for U-Net segmentation.

---

� Architecture Overview

Pipeline Diagram

graph TD;
    A[Raw CT Scans] --> B[Preprocessing & Normalization]
    B --> C[Patch Extraction & Dataset Prep]
    C --> D[U-Net Model for Segmentation]
    D --> E[Mask Extraction]
    E --> F[ROI Selection]
    F --> G[Candidate Generation for FPR]
    G --> H[CNN Training for False Positive Reduction]
    H --> I[Final Evaluation]

Loading

---

Notebooks Breakdown

Step	Notebook	Description
1	01_prepro.ipynb	Initial CT scan preprocessing
2	01_data_prep.ipynb	Patch extraction and formatting
3	01_data_analysis.ipynb	Exploratory Data Analysis
4	01_unet_train.ipynb	U-Net training for segmentation
5	02_maskextract.ipynb	Extracting binary masks
6	03_ROI.ipynb	Selecting regions of interest
7	01_FPR_candidate_roi.ipynb	Generating FPR candidates
8	02_FPR_dataset.ipynb	Building dataset for CNN
9	03_FPR_TRAIN.ipynb	CNN model training
10	04_FPR_CNN_eval.ipynb	Evaluating FPR model
11	LND_FINALE.ipynb	Final evaluation of pipeline

---

Evaluation Metrics

• Segmentation: Dice Score, Intersection-over-Union (IoU)
• Classification (FPR): Accuracy, Precision, Recall, F1-score, Confusion Matrix

---

Installation

git clone https://github.com/Anshul-ydv/LungNoduleSegmentationAndDetection.git
cd LungNoduleSegmentationAndDetection
pip install -r requirements.txt

---

How to Run

Follow the notebooks in sequence:

1. Start with 01_prepro.ipynb to preprocess CT scan data
2. Train U-Net using 01_unet_train.ipynb
3. Extract masks and define ROIs
4. Generate candidate ROIs and train CNN for FPR
5. Final evaluation in LND_FINALE.ipynb

---

Results Summary

Metric	U-Net Dice Score	CNN Accuracy (FPR)
Value	~0.82	~0.91

---

Citation

If you use this repository for your research or project, please cite:

@misc{lungnodule2025,
  title={Lung Nodule Detection and False Positive Reduction using Deep Learning},
  author={Anshul Yadav},
  year={2025},
  howpublished={\url{https://github.com/Anshul-ydv/LungNoduleSegmentationAndDetection}},
  note={GitHub Repository}
}

---

Contact

For issues, suggestions, or collaboration, feel free to open an issue or contact via GitHub. Or email - [email protected]

---