Deep Learning Approach for Early Stage Lung Cancer Detection

The paper addresses the challenge of late diagnosis in lung cancer, a leading cause of cancer mortality. It outlines cancer staging and emphasizes that earlier detection improves survival outcomes. CT scans are highlighted as the recommended imaging modality for lung cancer detection. The study aims to develop a deep learning model based on convolutional neural networks (CNNs) to predict and detect lung cancer and distinguish stages (benign, malignant, normal) from CT images. The motivation is to provide a tool that supports radiologists in early diagnosis, potentially reducing missed nodules and false positives and improving decision-making and treatment planning.

The related work surveys deep learning applications in lung cancer imaging. Jiang et al. (2022) used U-Net and 3D CNNs on DICOM and MHD formats to reduce false positives with high accuracy. Ahmed et al. (2020) employed a 3D CNN on the LUNA dataset with two convolutional and two max-pooling layers, achieving 80% accuracy on 400 test images. Rajan et al. (2020) proposed LungNet, a 53-layer hybrid CNN, evaluated on 19,089 CT images, reporting 0% false positives versus 6.4% for AlexNet. Mehta et al. (2021) combined 3D CNNs with Random Forests on LIDC-IDRI for benign/malignant classification; their biomarker-only Random Forest outperformed the hybrid model. Other studies include CNN+RNN frameworks for NSCLC survival and prognosis prediction, ResNet34 for detection (Yoo et al., 2020), and broader efforts such as Kirienko et al. (2018), Tekade and Rajeswari (2018), Hosny et al. (2018), Ardila et al. (2019), She et al. (2020), Song et al. (2017), Teramoto et al. (2016), Van Ginneken et al. (2015), and Zhang et al. (2019), covering tasks from nodule detection and malignancy estimation to survival prediction and treatment response. Comparative results on the IQ-OTH/NCCD dataset report accuracies in the 89–95% range for SVM/GoogLeNet/AlexNet baselines, providing a benchmark for the proposed method.

A CNN-based approach was developed to classify CT scans into benign, malignant, and normal. Model architecture: multiple convolutional layers with ReLU activation for nonlinearity, max-pooling layers to reduce overfitting, and a SoftMax output layer for three-class classification. Dataset: IQ-OTH/NCCD-Lung Cancer Dataset (Kaggle), totaling 1097 CT images across three classes: benign (15 cases; 120 images), malignant (40 cases; 561 images), and normal (55 cases; 416 images). Data preprocessing: images resized to 224×224. Original sizes included 512×512 for most, with a few images at 404×511, 512×62, 512×801, and 331×506. Contrast enhancement used CLAHE (OpenCV default tile size 8×8; clip limit 0.01; 256 histogram bins) with bilinear interpolation to avoid tile artifacts, followed by median filtering to reduce noise while preserving edges. Data augmentation: vertical flip (40%), horizontal flip (30% and additional horizontal flip via generator), random brightness (30% probability; min factor 0.3; max 1.2), random zoom (20%), rotation range 40°, shear range 20%, rescaling by 1/255, and height shift range 20%. Data split: 70% training, 15% validation, 15% testing. Implementation details: Python 3 with TensorFlow-Keras; training conducted on Google Colab Pro with Tesla P100 16 GB GPU and 32 GB RAM. Training ran for 10 epochs. After augmentation, the dataset expanded to 8461 images: 6345 for training (benign 961, malignant 3067, normal 2317) and 2116 for validation (benign 321, malignant 1023, normal 772).

• After augmentation, training/validation sets contained 6345/2116 images, respectively, across three classes.
• Confusion-matrix-derived metrics (on validation/testing):
  • Precision: benign 98%, malignant 1.00, normal 98%.
  • Recall (sensitivity): benign 98%, malignant 99%, normal 99%.
  • Specificity: benign 99.55%, malignant 99.45%, normal 98.65%.
  • F1-score: benign 0.98, malignant 1.00, normal 0.99.
  • Macro and weighted averages for precision, recall, and F1-score: 99%.
  • Overall accuracy: 99.45%.
  • Loss: 1.75%.
• Comparative results on the IQ-OTH/NCCD dataset show the proposed model outperforming prior CNN/SVM baselines (e.g., AlexNet ~93.45% accuracy; others ~89–95%) with sensitivity 99% and specificity 99.21% at 10 epochs.

The study demonstrates that a carefully designed CNN with contrast enhancement (CLAHE), noise reduction, and extensive augmentation can accurately classify CT images into benign, malignant, and normal categories. The high precision, recall, and specificity across classes indicate effective reduction of false positives and false negatives, directly addressing the need for early, reliable detection. Compared with prior models on the same dataset, the proposed approach achieves superior accuracy and balanced performance, suggesting it can serve as a supportive tool for radiologists to improve early-stage detection and staging decisions. The strong results on augmented data highlight the benefits of preprocessing and augmentation for handling real-world variability in CT imaging.

A CNN-based model was proposed for early prediction and detection of lung cancer from CT scans, classifying benign, malignant, and normal cases. The method achieved 99.45% accuracy with strong precision, recall, specificity, and F1-scores, and reduced false positives. Given the critical importance of early detection for improving survival, this model can support radiologists’ diagnostic decisions. Future work could include external validation on larger, multi-institutional datasets, testing generalizability across scanners/protocols, integrating explainability, and extending to end-to-end nodule detection and staging pipelines.