PSO-MCAE-LCCD: Particle Swarm Optimized MobileNetV2 Convolutional Autoencoder for Lung and Colon Cancer Detection

Abstract

Identifying lung and colon cancer at an early stage significantly increases their chances of survival. This research proposes a new deep learning model PSO-MCAE-LCCD, which combines feature extraction using MobileNetV2 with classification using a Convolutional Autoencoder (CAE) implemented with Particle Swarm Optimization (PSO). To improve the median filtering, histopathological images from the LC25000 dataset were pre-processed. PSO optimally tunes the learning rate, dropout rate, and unit dense ratio, which significantly enhances model performance. After testing with 80:20 and 70:30 train-test splits, the model achieved high accuracy, precision, recall, and F1-score for all five cancer classes. Proposed model validation results showed 99.38% accuracy, outperforming other models in computation with a prediction time of 17.95 seconds. ROC and precision-recall curves validate model performance for all tested classes. Results show that PSO-MCAE-LCCD is a robust and efficient tool for automated histopathological cancer detection.