Optimizing digital histologic colorectal cancer detection using MobileNet-based transfer learning

Purpose: Colorectal cancer (CRC) is a major global health challenge, with an increasing incidence among younger populations. However, traditional screening methods lack comprehensiveness. The proposed work aimed to develop and evaluate a lightweight, accurate, deep learning model for classifying CRC from histo-logical images using MobileNetV3 (Google Research) -based transfer learning. Methods: This study presents a MobileNetV3-based transfer learning model, trained and validated on two publicly available datasets: LC25000 and Katherₜexture₂016. The model was fine-tuned using optimized hyperparameters and evaluated in a Python-based environment with graphics processing unit (GPU) support. The performance metrics included classification accuracy and latency. Results: The proposed MobileNetV3-based model demonstrated high classification accuracy across all cat-egories and exhibited robust performance, even for cancer types not seen during training. The model achieved an average detection latency of approximately 0. 2014 seconds per sample. These results highlight the efficiency of the model and its potential for integration into the clinical workflow. Conclusion: The proposed MobileNetV3-based transfer learning model offers a scalable and effective solution for analyzing CRC histological images. While the performance on benchmark datasets is promising, an external test using real-world clinical data is needed to support broader clinical deployment. Future studies will focus on external testing using hospital-grade datasets and on expanding the model’s capabilities to other cancer types.