A robust transfer learning approach for colorectal cancer identification based on histopathology images

Diagnosis of cancer at the benign stage is crucial. Recently, pathologists have been using computer-aided diagnostics with machine learning to diagnose patients from medical images. The limitation of the medical image dataset is a challenge to obtain a robust model for cancer identification. This research was conducted to address this issue. The weight of the pre-trained model can be learned more efficiently with a limited amount of data. However, the training process is conducted by customizing the freeze rate so that feature extraction is preserved besides tuning some hyperparameters.

Transfer learning is a technique that can handle data limitation issues in the medical field. Transfer learning will reuse pre-trained model for different or specific task. Choosing an optimum architecture and hyperparameters in machine learning is very important to improve model performance. In our experiment, we carried out a hyperparameter optimization of various deep learning architectures that classifies images containing healthy and cancer tissue.

The research concludes that CNN with architecture DenseNet121, freeze rate 75%, zero hidden layers on the classifier, learning rate 0.001, and optimizer RMSProp have the best performance with 98% accuracy and 19.5 s training time using NVIDIA A100 GPU accelerator. Testing with the real dataset for future direction will be an achievement for the model’s success.

This research has successfully optimized the Densnet121 deep learning architecture by tuning parameters. With a harmonic means value of 0.98, DenseNet121 outperforms the others. Compared to Camelyon, ImageNet still becomes the baseline of the transfer learning dataset because it has a rich amount of data.