Abstract:
Accurate tumor staging is essential for optimizing colorectal cancer treatment planning and improving patient outcomes. The interpretation of histopathological slides is used as the foundation of cancer staging, in the modern practice. However, inter-observer variability can also undermine this method and moreover there is the complex morphology of tumours. Computational pathology and the development of multi-omics data have given a leading opportunity to improve the accuracy of the staging through the combination of heterogeneous modalities. In this dissertation, we are going to propose and evaluate a multimodal framework with fusion of whole-slide image (WSI) features, mRNA expression features, and clinical metadata (age) in order to do automated classification of colorectal cancer T-stage (T1-T4). A compound synthesizing dataset was constructed, including 1024 features derived on the basis of images, 128 mRNA features and the age data on 261 patients. After cleaning, preprocessing, and stratified partitioning, the normalized resultant feature vectors were used to train two families of predictive models: a lightweight convolutional neural network called TLCNN 2 used on large advise samples sizes with high-dimensional tabular fusion data, and an XGBoost classifier, which is especially effective on the small-sample space with high-dimensional biomedical data. The TLCNN-v2 achieved high results on the training data but was affected by over-fitting, suggesting that this was based on the small size of the training cohort. On the other hand, the XGBoost-Fusion model provided better generalisation and had the high final test accuracy of 57 per cent of the four tumour-stage classes and outdid other methods, including MLPs, PCA-reduced models, convolutional-net-based models, and augmented-WSI chunking strategies.Analytical results indicate that multimodal fusion improves discriminative power, and XGBoost provides stable learning under class imbalance and high feature dimensionality. This work demonstrates that integrating histopathology, molecular expression, and clinical indicators meaningfully improves automated colorectal cancer staging. Future extensions may incorporate patch-level WSI transformers, multi-omics attention models, and data-level augmentation to further enhance performance.
