Detection, localization, segmentation, and classification in colorectal cancer screening using deep learning: A systematic review

Abstract

Colorectal polyps can develop into colorectal cancer (CRC), one of the leading causes of cancer-related deaths. These polyps must be found and treated as soon as possible to avoid developing CRC. Artificial intelligence (AI), especially deep learning (DL), has markedly improved the early diagnosis and treatment of CRC by increasing the accuracy and efficiency of polyp identification and classification. However, the rapid growth of research in this area has created a fragmented environment, highlighting the need for a comprehensive synthesis of findings to guide future advancements. This review aims to address this gap by systematically analyzing the state-of-the-art DL methodologies applied to colorectal polyp analysis. Adhering to the Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA) guidelines, we conducted a structured search of major databases, identifying and analyzing 160 papers published between 2017 and 2024. The review focuses on the architectures of DL models, publicly available datasets, and performance metrics. The authors have analysed a wide range of DL architectures, including Convolutional Neural Networks (CNNs), YOLO-based object detectors, transformer models, recurrent neural networks (RNNs), autoencoders, and hybrid systems. Notably, YOLOv4 and CA-ResNet50 displayed state-of-the-art detection performance with accuracy rates exceeding 99 % on benchmark datasets such as LC25000 and Kvasir-SEG. For segmentation tasks, transformer-enhanced models like ViT and SwinE-Net earned excellent Dice scores (>0.92), beating standard UNet variations. However, the research also reveals persisting problems, such as dataset variability, low generalizability across diverse populations, and the absence of standardized benchmarking techniques. In addition, we critically investigate the significance of data augmentation strategies in reducing dataset limits and overfitting, and also analyze how dataset characteristics affect model performance. Finally, structured future recommendations have been provided across three critical dimensions: (1) designing comprehensive and diverse datasets to improve generalizability, (2) establishing standardized and task-specific data augmentation pipelines, and (3) developing hybrid and modular architectures optimized for specific diagnostic subtasks. This thorough analysis aims to clarify the present status of research and highlight possibilities for enhancement in this vital healthcare sector.