Deep learning in lower gastrointestinal cancer detection: Advances in endoscopic, radiologic, and histopathologic diagnostics

Table 1 Overview of artificial intelligence technologies applied in medical research and clinical practice

AI technology	Type	Typical applications	Data modality	Examples
Convolutional neural networks[9]	Deep learning	Image classification, lesion detection, segmentation	Endoscopy, CT, MRI, ultrasound, histology	Polyp detection in colonoscopy, tumor segmentation on CT, WSI classification
Deep learning (radiomics/segmentation models)[17,27]	Deep learning	Temporal data analysis, prognosis prediction	Longitudinal EMR data, time-series labs	Predicting recurrence or survival in CRC or HCC
Transformers[21]	Deep learning (NLP)	EHR analysis, pathology report interpretation, literature mining	Text (EMRs, reports)	Extracting staging from pathology notes, summarizing clinical texts
Support vector machines[25]	Machine learning (supervised)	Classification tasks, small datasets	Genomics, proteomics, structured data	Classifying tumor vs normal tissues based on gene expression
Random forests/decision trees[39]	Machine learning (supervised)	Risk prediction, survival modeling	Clinical + demographic data	CRC risk prediction based on age, lifestyle, family history
Multimodal deep learning models (e.g., MuMo)[44]	Deep learning/multimodal AI	Predicting treatment response, prognostic modeling; survival prediction	Radiology (CT/MRI), pathology images, genomic/transcriptomic data, clinical variables	MuMo integrates diverse data types (radiological, pathological, and clinical information) to predict treatment responses and survival outcomes for patients with HER2-positive GC
Explainable AI[47]	Deep learning/explainable AI	Enhancing clinician interpretability and trust; feature attribution; transparent model reasoning	Histopathology, endoscopic images	PolypSeg-GradCAM integrates U-Net segmentation with Grad-CAM visualization for pixel-level explanation in colonoscopy; pathology explainable AI identified histological features linked to molecular subtypes or therapy response
Federated learning models[49,50]	Machine learning/federated learning	Multicenter model training without sharing raw patient data; improving generalizability and equity	Multi-institution clinical, imaging, and omics data	Robust federated learning model for predicting postoperative recurrence in GC outperformed clinical model and reduced misdiagnosis of local recurrence by > 40%
Digital twin technology[51]	Hybrid AI/simulation	Virtual patient modeling for disease trajectory simulation, treatment response, and outcome forecasting	Longitudinal clinical, imaging, and molecular data	Simulating tumor evolution and therapy outcomes for colorectal and GCs; currently in research validation stages

AI: Artificial intelligence; CRC: Colorectal carcinoma; EHR: Electronic health record; GC: Gastric cancer; HCC: Hepatocellular carcinoma; MRI: Magnetic resonance imaging; NLP: Natural language processing; WSI: Whole slide image; CT: Computed tomography; HER2: Human epidermal growth factor receptor 2; EMR: Electronic medical record.