Complex or Irregular Lung Nodule Detection from CT Scans
Abstract
Lung cancer is a leading cause of cancer-related deaths, with a low survival rate. Early detection using medical imaging is crucial. However, CT scans often fail to detect irregular or complex nodules. This project leverages deep learning V-Net, YOLOv5s, and DETR models combined with a shape modeling algorithm to accurately detect lung nodules and reduce misinterpretation.
Problem Statement
Detecting lung nodules manually in CT scans is time-consuming and prone to errors. Nodules vary in size, shape, and location, which leads to high false-positive rates. Automated detection can improve radiologist efficiency, assist treatment planning, and enable remote diagnostics.
Applications
- Research & Clinical Studies: Evaluate new treatments.
- Treatment Planning: Personalize care based on nodule growth and location.
- Radiologist Assistance: Improve workflow by handling routine detection tasks.
- Telemedicine: Enable remote, timely consultations.
- AI Integration: Combine with other AI models for comprehensive diagnosis.
Challenges
- Variability in nodule characteristics (size, shape, density).
- High false positives causing unnecessary tests.
- Adapting models to new datasets from different scanners or hospitals.
Proposed System
We designed a multi-model system:
V-Net
A 3D convolutional neural network for volumetric segmentation.
- Accurately segments lung nodules of varying shapes and sizes.
- Uses residual functions and weighted learning for complex patterns.
- Input: CT scan images with corresponding masks.
- Output: Precise nodule segmentation.
YOLOv5s
Real-time object detection model for nodule localization.
- Processes each CT slice as an individual image.
- Assigns probabilities to detected nodules in N×N grids.
- Integrates with V-Net for improved detection accuracy.
YOLOv5s + Shape Modeling Algorithm
- Captures local shape features: compactness, eccentricity, solidity.
- Helps classify nodules as solid or ground-glass opacity.
- Reduces false positives by analyzing shape and intensity.
DETR (Detection Transformer)
- Transformer-based model for refined object detection.
- Uses queries to detect nodules in images.
- Outputs class labels and bounding boxes using a Multi-Layer Perceptron (MLP).
Implementation
Data Preparation
- Resizing: Images resized to 256×256 pixels to balance detail and computational efficiency.
- Diagnosis Mapping: Masks analyzed to mark presence of nodules (1 if present, 0 otherwise).
- Data Splitting: Train, validation, and test sets created using
train_test_split. - Data Augmentation: Rotations, flips, shear, zoom applied to improve generalization.
Model Training
- V-Net: Encoder-decoder architecture with convolutional blocks, batch normalization, ReLU activations, and fully connected layers for feature extraction.
- YOLOv5s: Single-pass object detection for bounding box prediction.
- DETR: Transformer encoder-decoder for enhanced detection and localization.
Results
Sidenote: Sample detections by YOLOv5s- V-Net: High accuracy in nodule segmentation using Dice coefficient and IoU metrics.
- YOLOv5s: Accurate detection of multiple nodules per CT slice, real-time inference.
- DETR: Refined detection, reducing false positives.
Conclusion
The multi-model approach combining V-Net, YOLOv5s, DETR, and shape modeling successfully detects complex lung nodules from CT scans. This system improves early detection, reduces misinterpretation, and supports radiologists in clinical settings.
Future Enhancements
- Integrate with clinical workflow for real-time detection.
- Expand dataset for better generalization across hospitals.
- Explore advanced transformers for improved feature representation.
Tools & Libraries
- Python, TensorFlow, Keras
- Scikit-learn, Numpy, Matplotlib
- Ultralytics YOLOv5s package
Dataset
- Lung Image Database Consortium (LIDC-IDRI)
- CT scans with annotated lung nodules
View PDF