TWO-STAGE FINE-TUNING OF PRETRAINED CNN MODELS FOR MULTICLASS LUNG CANCER CLASSIFICATION USING CT SCANS

ABSTRACT

Detection of lung cancer at its early stage using CT images can significantly enhance diagnosis and treatment outcomes. This study implements a two-phase approach to fine-tune five pretrained CNN models – VGG19, ResNet50, InceptionV3, DenseNet121, and EfficientNetB0 – for the classification of lung CT scans into benign, malignant, and normal categories. In Stage 1, models were adapted using benign and normal images, and then fine-tuned in Stage 2 for three-class classification. These models were evaluated using metrics such as accuracy, sensitivity, specificity, and AUC. InceptionV3 surpassed the rest with 98.8% accuracy and 0.999 AUC, followed closely by ResNet50. While VGG19 exhibited lower classification accuracy and sensitivity, DenseNet121 and EfficientNetB0 demonstrated strong performance. This study highlights the effectiveness of transfer learning and domain-specific fine-tuning in medical imaging, identifying InceptionV3 and ResNet50 as the most robust architectures for future development of predictive models for early lung cancer risk assessment.

Keywords: Fine-tuning, pretrained CNN, Computed Tomography, Transfer learning