Optimized Deep Learning Architecture for Early Detection of Lung Cancer

Abstract

Lung cancer has been among the causes of cancer related deaths worldwide. Computed tomography (CT) imaging can be used to detect the disease at an early and accurate stage to enhance patient survival rates. Nevertheless, there are inter-observer variability, time constraints and cognitive fatigue with the manual radiological interpretation. The research provides a recent deep learning-based automated lung cancer classification based on CT scans with an EfficientNetV2 architecture. The given system was trained and tested on a large dataset of 315 test samples that were classified into four clinically important groups adenocarcinoma (T2N2M0IIa), large cell carcinoma (T2N2M0IIia), normal lung tissue, and squamous cell carcinoma (T1N2M0_IIa). The EfficientNetV2-M model has excellent performance scores: test accuracy of 94.92, macro-averaged precision of 94.64, recall of 95.57, and F1-score of 95.03. It is a great enhancement of 2.92 percentage points of the best accuracy of 92 that was earlier reported in the literature [6]. The model itself uses the latest training methods such as label smoothing, class-weighted loss functions, mixed-precision training, cosine annealing with warm restarts, and progressive data augmentation in medical imaging in particular. The convergence nature of the training process has excellent converging properties with low overfitting properties as shown by the train-validation gap that has consistently constituted during the 100 epochs of the training. The results can be compared with ResNet18 (92.06% accuracy) and DenseNet121 (92.70% accuracy), which proves that EfficientNetV2 is superior in this case of the classification. There are outstanding results of discrimination between normal and pathological cases as revealed by per-class analysis of 98.15% recall of normal tissue and 96.08% recall of large cell carcinoma and a robust 92.50% recall of adenocarcinoma classification.