An optimized cascaded transformer with progressive attention for lung and colon cancer diagnosis from histopathological images.

Lung and colon cancers are the most fatal diseases universal, necessitating early and accurate diagnosis. Traditional deep learning models frequently face disadvantages such as overfitting, high computational cost, and limited scalability. To address these, a novel transformer-based framework is proposed for histopathological image classification. Initially, images are pre-processed using a Gaussian Adaptive Smooth Kalman filter (GASK) to enhance quality. Feature extraction is performed using a Convolutional Inception Enclosed MobileNet (CIEM), which integrates inception layers for precise pattern recognition. The classification is handled by a Progressive Attention-based Optimized Cascaded Transformer Network (PCTN), enabling efficient patch-wise analysis. Hyperparameters are fine-tuned using the Random Chaotic Bald Eagle Optimization algorithm. The proposed model achieves 98.9% and 98.57% accuracy on lung and colon cancer datasets, respectively, outperforming existing methods in precision, sensitivity, and F1-score.