TL-LASSO-Net: a hybrid transfer learning and LASSO-based framework for robust colon cancer histopathology classification on LC25000 and GlaS datasets.

Sorting colorectal histology images helps physicians diagnose then cure. Deep CNNs excel in digital pathology. However, their high-dimensional feature spaces cause overfitting, interpretability issues, and decreased generalization across datasets with different staining methods than glandular structures. The hybrid system TL-LASSO-Net combines transfer learning-driven deep feature extraction with LASSO regression for better sparse feature selection. To get high-level features from pre-trained backbones (ResNet50, DenseNet121, and ViT-B/16 for comparison) and compress them using logistic LASSO. After that, a lightweight dense classifier works in a smaller, more discriminative subspace. Two benchmark histopathology datasets, LC25000 and GlaS, are used for the experiments. TL-LASSO-Net gets 98.3% accuracy and ROC-AUC of 0.997 on LC25000, which is better than ResNet50-TL (97.5%, AUC 0.993) and DenseNet121-TL (97.3%, AUC 0.992). The suggested method gets 97.2% accuracy and 0.994 ROC-AUC on GlaS, which is better than all the other methods. Cross-dataset evaluations (LC25000→GlaS and GlaS→LC25000) further show that generalization has gotten better. TL-LASSO-Net has an accuracy of up to 91.3%, whereas the best baseline has an accuracy of just 88.9%. The results show that combining transfer learning with LASSO-driven sparsity makes a model that is small, easy to understand, and quick to compute that can be used to help computers diagnose colorectal cancer.