Deep Learning and Optimization Approaches in DeepLabV3-DenseNet: Leveraging Radiomics Feature Extraction and Non-Invasive Detection of Microsatellite Instability in Colorectal Cancer with a Hyperparameters-Tuned Pre-trained Model: A Review

Abstract

Microsatellite instability (MSI) is a critical biomarker in colorectal cancer (CRC), influencing prognosis and response to immunotherapy. Traditional MSI detection methods rely on invasive biopsy and molecular testing, which are time-consuming and costly. Recent advancements in deep learning and radiomics have enabled non-invasive MSI prediction using medical imaging modalities such as CT, MRI, and histopathology slides. In particular, convolutional neural networks such as DeepLabV3 and dense Net have demonstrated superior performance in feature extraction and segmentation tasks. Radiomics-based approaches extract high-dimensional quantitative features from medical images, capturing tumour heterogeneity and underlying biological characteristics. Studies have shown that machine learning models using PET/CT radiomics can effectively predict MSI status with high accuracy. Furthermore, deep learning-based frameworks trained on histopathology images have achieved strong predictive performance with AUC values exceeding 0.9. Hybrid architectures combining DeepLabV3 for segmentation and dense Net for classification further enhance feature representation and model performance. This review explores recent advancements in deep learning and optimization-based MSI detection, focusing on radiomics, graph-based learning, and hyperparameter optimization. The study highlights the advantages of hybrid deep learning models and discusses current challenges, including data dependency, interpretability, and computational complexity.