Fully automated tumor-stroma ratio prediction as prognosis factor in colorectal cancer using optimized deep transfer learning

The Tumor-Stroma Ratio (TSR) reflects a higher ratio of stromal cells compared to cancer cells in a tumor, which is associated with a worse prognosis, increased risk of recurrence, and reduced overall survival in colorectal cancer. This paper reports the application of optimized deep transfer learning for fully automated tumor-stroma ratio estimation and its significance as a prognostic factor in patient survival prediction. Various base models of transfer learning, including ResNet18, DenseNet201, MobileNet_V2, EfficientNet_b2, and Inception_V3, are considered. All base models were trained using the Optuna hyperparameter optimization framework, which is flexible, modular, and easy to combine with transfer learning. After determining the optimal deep transfer learning model, it was implemented to predict the tumor-stroma ratio (TSR). Experimental comparisons indicate that ResNet18 is the most suitable base deep transfer learning model for colorectal cancer (CRC) classification. Consequently, ResNet18 was used to classify whole slide images of colorectal patients. Finally, the Cox proportional hazards model was applied to features including TSR and other clinical parameters of colorectal cancer. The results demonstrate that the Tumor-Stroma Ratio can be considered an important prognostic factor for colorectal cancer based on survival analysis prediction.