Automated tumor stroma ratio assessment in colorectal cancer using hybrid deep learning approach

The Tumor-Stroma Ratio (TSR) is a critical prognostic factor in colorectal cancer (CRC), offering insights into tumor microenvironment interactions. However, traditional TSR assessment methods are subjective and labor-intensive. This study is among the first approaches that introduce a novel integrative deep learning to leverage Transformer mechanisms for enhanced spatial context understanding in TSR assessment for CRC, addressing limitations in conventional CNN-only models and combining. Our methodology involves classifying patch images into normal and abnormal categories using a CNN-based model, then segmenting abnormal patches into tumor and stroma areas in whole-slide images (WSIs) through a hybrid CNN-Transformer UNET model. TSR is quantitatively assessed based on pixel area, providing a robust and objective measure of the tumor microenvironment. The classification model performed superiorly by achieving an overall classification accuracy of 93.53%, precision of 0.9040%, recall of 0.9644%, f1-score of 0.9332%, and Matthew's correlation coefficient (MCC) of 0.8724, sufficient to maintain consistent TSR evaluation in case of eventual misclassification. In segmentation tasks, the proposed Efficient-TransUNet model achieved the highest Aggregated Dice Coefficient (ADC) of 0.938 for stroma and 0.921 for tumor class, outperforming existing models. These results underscore the potential of this hybrid deep learning framework to enhance the precision of pathological evaluations and contribute to improved clinical outcomes.