Lung Nodule Detection and Classification from Thorax CT Scans using RetinaNet with Transfer Learning

Abstract

This research investigates the application of RetinaNet, a one-stage object detection model, coupled with transfer learning, for the automated detection and classification of lung nodules in thoracic CT scans. The objective is to improve the accuracy and efficiency of lung nodule identification compared to existing methods. A pre-trained ResNet backbone is fine-tuned on a dataset of annotated CT scans, leveraging transfer learning to mitigate the need for extensive labeled data. The results demonstrate improved performance metrics, particularly in terms of precision and recall, compared to baseline methods. This approach promises to aid radiologists in early diagnosis and improved patient outcomes.

Introduction

Lung cancer is a leading cause of cancer-related deaths globally, and early detection is crucial for successful treatment. Thoracic CT scans are a primary diagnostic tool, but manual analysis is time-consuming and prone to inter-observer variability. Automated lung nodule detection and classification systems are essential for improving diagnostic accuracy and efficiency. However, challenges remain in achieving high sensitivity and specificity, particularly in distinguishing malignant from benign nodules, due to the subtle variations in nodule appearance and the presence of artifacts in CT scans. This project aims to address these challenges using a deep learning-based approach.

Objectives

• To develop an accurate and efficient lung nodule detection system using RetinaNet with transfer learning.
• To improve the classification accuracy of lung nodules into benign and malignant categories.
• To evaluate the performance of the proposed system using standard metrics like precision, recall, F1-score, and AUC.