[Paper Review] Deep Learning Approach for Early Stage Lung Cancer Detection

Saleh Abunajm, Nelly Elsayed+2 more|arXiv (Cornell University)|Feb 5, 2023

Lung Cancer Diagnosis and Treatment9 citations

TL;DR

The paper presents a CNN-based model for early prediction and detection of lung cancer stages from CT scans, achieving high accuracy and aiming to support radiologists in diagnostic decisions.

ABSTRACT

Lung cancer is the leading cause of death among different types of cancers. Every year, the lives lost due to lung cancer exceed those lost to pancreatic, breast, and prostate cancer combined. The survival rate for lung cancer patients is very low compared to other cancer patients due to late diagnostics. Thus, early lung cancer diagnostics is crucial for patients to receive early treatments, increasing the survival rate or even becoming cancer-free. This paper proposed a deep-learning model for early lung cancer prediction and diagnosis from Computed Tomography (CT) scans. The proposed mode achieves high accuracy. In addition, it can be a beneficial tool to support radiologists' decisions in predicting and detecting lung cancer and its stage.

Motivation & Objective

• Motivate early lung cancer diagnostics to improve survival rates.
• Develop a deep learning model to predict and detect lung cancer stages from CT images.
• Evaluate the model against established metrics and compare with state-of-the-art methods.

Proposed method

• Propose a deep convolutional neural network with multiple Conv2D and MaxPooling layers followed by dense SoftMax for three-class output (benign, malignant, normal).
• Use ReLU activations and early pooling to prevent overfitting.
• Train on the IQ-OTH/NCCD-Lung Cancer Dataset with image preprocessing and CLAHE-based contrast enhancement.
• Apply data augmentation to expand the dataset (flips, brightness, zoom, rotation, shear, etc.).
• Report detailed metrics including precision, recall, specificity, F1-score, and overall accuracy.

Experimental results

Research questions

• RQ1 Can a CNN-based model accurately classify CT scans into benign, malignant, and normal categories for early lung cancer detection?
• RQ2 What is the impact of CLAHE-based contrast enhancement and data augmentation on model performance?
• RQ3 How does the proposed model compare to state-of-the-art CNN models on the IQ-OTH/NCCD dataset?

Key findings

• The model achieved an accuracy of 99.45% on the augmented IQ-OTH/NCCD dataset.
• Precision, recall (sensitivity), and specificity for the three classes are high, with malignant precision at 1.00 and recall of 99%.
• F1-scores are 0.98 for benign, 1.00 for malignant, and 0.99 for normal cases.
• The macro and weighted averages for precision, recall, and F1-score are 99%.
• Training used 6,345 augmented images and 2,116 validation images, yielding a loss of 1.75.