Lung Cancer Diagnosis: Small Lung Nodule Dataset

This project builds upon the Small Lung Nodule Lung Cancer Dataset to classify small lung nodules, aiming to enhance early diagnosis of lung cancer. Our approach outperforms the baseline performance outlined in the referenced manuscript in several aspects, offering a robust decision-support tool for radiologists.

---

Dataset Overview

This dataset accompanies the manuscript:
"Radiomics-Based Decision Support Tool Assists Radiologists in Small Lung Nodule Classification and Improves Lung Cancer Early Diagnosis."

• Dataset Link: Small Lung Nodule Dataset

The dataset includes 990 nodules from 810 patients, annotated with features extracted from multi-region CT segmentations using TexLab2.0. It introduces a predictive radiomics model, Small Nodule Radiomics Predictive Vector (SN-RPV), developed through LASSO regression for nodule classification.

---

Objectives

• Enhance Radiologist Evaluation: Support radiologists in lung nodule classification and improve early cancer detection.
• Risk Stratification: Categorize lung nodules into Low Risk, Intermediate Risk, and High Risk groups.
• Quantifiable Improvements: Offer significant enhancements in missed or delayed cancer diagnoses compared to radiologists.

---

Workflow

Radiologist Evaluation

1. Initial Assessment: Radiologists evaluate lung nodules using CT scans.
2. Risk Categorization: Nodules are stratified into three categories based on SN-RPV:
   • Low Risk: Unlikely to be cancerous.
   • Intermediate Risk: Moderate likelihood of malignancy.
   • High Risk: High probability of cancer.

Management Decisions

• Low Risk:
  • SN-RPV High → Place patient in a "Safety Net" with no further follow-up.
• Intermediate Risk:
  • SN-RPV High → Early surveillance CT to monitor growth and enable early diagnosis.
• High Risk:
  • Aggressive investigation, including biopsy or advanced imaging.

---

Results and Improvements

Quantifiable Improvements

Our approach outperforms the referenced manuscript in several key metrics:

• Improved AUC for malignancy prediction across multiple methods and models.
• Enhanced accuracy and sensitivity for detecting small lung nodules.
• Achieved 73% accuracy and a 66.67% improvement in missed or delayed cancer diagnoses compared to the average radiologist.

Model Performance

Feature Selection Method	Model	Accuracy	Sensitivity	Specificity	AUC
Boruta	SVC	0.7323	0.8056	0.6364	0.7775
Boruta	Neural Network	0.6850	0.6944	0.6727	0.7636
Mutual Information	Neural Network	0.7244	0.7639	0.6727	0.7662
XGBoost	Neural Network	0.7323	0.8056	0.6364	0.7619

Comparison with Radiologists

The test set's 73% accuracy surpassed the average performance of six thoracic radiologists, offering earlier and more reliable cancer detection.

---

Future Directions

• Expand Validation: Incorporate external datasets for broader evaluation.
• Deep Learning: Explore CNNs for feature extraction and classification.
• Personalized Decision Tools: Enhance decision-support tools for tailored patient management.