Cancer Genetics and Epigenetics 2024, Vol.12, No.4, 210-222 http://medscipublisher.com/index.php/cge 214 Figure 2 An illustration of the neural network structure (Adopted from Xiao et al., 2018) Image caption: This diagram illustrates the structure of an artificial neural network, including the input layer, hidden layers, and output layer. The input layer is responsible for receiving external input data. The hidden layers consist of multiple nodes and are responsible for performing complex nonlinear transformations on the input data. Each node in the hidden layers is connected to all nodes in the previous layer. The multi-layer connections between hidden layers enable the network to learn and represent complex patterns and features. Finally, the output layer generates the final prediction or classification result. Through the training process, the network adjusts the connection weights between layers to improve prediction accuracy. This structure is widely used in machine learning and deep learning for various classification and regression tasks (Adopted from Xiao et al., 2018) 4.2.1 Convolutional neural networks (CNNs) Convolutional Neural Networks (CNNs) are particularly effective in processing and analyzing image data. In colon cancer prediction, CNNs have been used to identify histopathological features from biopsy images, achieving high accuracy in distinguishing between cancerous and non-cancerous tissues. For instance, a study proposed a multi-scale feature fusion CNN based on shearlet transform, which achieved an identification accuracy of 96% for colorectal adenocarcinoma epithelium and normal colon mucosa (Liang et al., 2020). Another study developed a CNN model that integrates multiparametric MR-US imaging data and fusion biopsy trajectory-proven pathology data for 3D prediction of prostate cancer, demonstrating the potential of CNNs in multi-modal data fusion (Kaneko et al., 2022). 4.2.2 Recurrent neural networks (RNNs) Recurrent Neural Networks (RNNs) are designed to handle sequential data and have been applied in the analysis of time-series data in medical research. Although less common in image analysis compared to CNNs, RNNs can be useful in integrating longitudinal patient data, such as follow-up records and treatment responses, to predict cancer progression and recurrence. For example, a study utilized a deep learning-based multi-model ensemble method that incorporated RNNs to improve cancer prediction accuracy by leveraging gene expression data (Xiao et al., 2018). 4.3 Statistical and computational methods In addition to deep learning, various statistical and computational methods play a crucial role in the preprocessing and integration of multi-modal data for AI-based colon cancer prediction. 4.3.1 Feature selection and dimensionality reduction Feature selection and dimensionality reduction techniques are essential for handling high-dimensional data and improving model performance. These methods help in identifying the most informative features from large datasets, thereby reducing computational complexity and enhancing prediction accuracy. A study on multi-task multi-modal learning for cancer diagnosis and prognosis emphasized the importance of selecting informative features from histopathological images and genomic data to improve prediction performance (Shao et al., 2020). Similarly, another study developed a non-invasive imaging prediction model for liver metastasis in colon cancer
RkJQdWJsaXNoZXIy MjQ4ODYzNQ==