Artificial intelligence and machine learning applications in abdominal tuberculosis diagnosis: A scoping review and translational roadmap

Table 1 Characteristics and performance metrics of artificial intelligence and machine learning 16 original studies in abdominal tuberculosis diagnosis

Ref.	Country	Design and sample	Target	Primary modality	AI/ML technique	Key performance	Explainability	Key notes
Park et al[17], 2025	South Korea	Retrospective; internal 1132 CD + 1045 GITB images (130 points + 123 points; train/value/test), external 67 CD + 63 GITB images (17 points + 14 points, no overlap). Colonoscopy (white-light lesion images)	ITB vs CD (colonoscopic)	Colonoscopy (white-light lesion images)	Deep learning/CNN UNet-ResNet50 CNN (multitask learning)	Internal: Acc 97.7% (AUROC = 0.997). External: Acc 81.5% (AUROC = 0.877)	Human crossover trial (6 endoscopists, 780 interpretations): Pooled acc ↑86.2% to 88.8% (P = 0.010), trainees ↑76.7% to 81.0% (P = 0.002)	Most rigorous image-based study with external validation + human-AI crossover trial; demonstrates trainee benefit (+4.3% Acc) and addresses prior gaps in raw-image AI, generalizability, and clinical utility evaluation
Cheng et al[18], 2024	China	Retrospective; 330 patients (value 1 = 224, value 2 = 106); SMOTE	CD vs ITB	CT enterography (arterial + venous phases)	DL radiomics + LASSO + logistic	Arterial-venous DL AUC = 0.885/0.877/0.800 (train/value 1/value 2)	None reported	Multi-phase fusion key; DL outperformed handcrafted in arterial phase
Shu et al[19], 2024	China	Retrospective; 241 patients (51 parameters); real-world validation	ITB vs CD	Clinical/EHR/laboratory (51 parameters)	XGBoost + SHAP + LIME	AUC = 0.946; Acc 0.884; real-world Acc 0.860, Sens 0.833, Spec 0.871	SHAP + LIME; top features T-SPOT, pulmonary TB	Strong MDT agreement (90.7%, kappa 0.78); ready for deployment
Lu et al[20], 2024	China	Retrospective; 82 patients (25 ITB, 57 CD); train 54, test 28	ITB vs CD	Multimodal (MRE radiomics + colonoscopy radiomics)	LASSO + DL + logistic; fused model (33 features)	Multidisciplinary AUC = 0.94 (test); single-modality 0.68-0.90	None reported	First true multidisciplinary fusion; superior to all singles (DeLong P < 0.05)
Liu et al[21], 2024	China	Retrospective; 85 cases (1973 WSI)	CD vs ITB (surgical specimens)	Histopathology whole-slide images	DL CNN; attention maps	Case-level AUC = 0.886 (train)/0.893 (test); slide-level 0.954/0.827	Attention maps; top-10 features	Outperformed juniors; comparable to senior GI pathologists
Li et al[22], 2024	China	Retrospective; 72 paraffin sections	CD vs ITB	ATR-FTIR spectroscopy (mid-IR on paraffin)	Spectroscopy + XGBoost/ML	High accuracy and specificity (detailed in full text)	None reported	Rapid chemical fingerprinting; non-destructive
Lin et al[23], 2024	China	Retrospective; 122 endoscopic images (99 CD, 23 ITB)	ITB vs CD	White-light endoscopic (ileum)	Few-shot (2-way 3-shot; Xception + dual transfer)	AUC = 0.81 (dual transfer); better than single transfer (0.56) & endoscopists	None reported	Highly data-efficient for rare ITB; dual transfer learning
Shen et al[24], 2024	China	Retrospective; 301 patients (107 ATB-LN + 194 lymphoma)	Nodal TB vs lymphoma	CECT (abdominal LNs, 3D VOI)	Radiomics (8 features); LR (primary)	LR AUC = 0.91 (train)/0.86 (value); Acc 88%/83%; Sens 76%/72%	None reported	Strong shape/texture signature; excluded small/necrotic nodes
Pang et al[25], 2023	China (multicenter)	Retrospective; 178 patients (88 PTB + 90 CD)	PTB vs PC	CT (omental, peritoneal, ascites, LN signs)	ML ensemble on selected features	AUC = 0.971 (train)/0.914 (test); F1 0.923/0.867	None reported	Multicentre; simple interpretable CT signs; strong performance
Gong et al[26], 2023	China	Retrospective; 105 patients (61 CD, 44 ITB); 5-fold cross-validation (single-center)	ITB vs CD	CT enterography (multiregional radiomics)	Multiregional radiomics model (CT enterography-based)	Combined (nomogram) model: AUC = 0.975 (train 95%CI: 0.953–0.998)/0.958 (CV 95%CI: 0.925–0.991); Acc 89.5%; Sens 86.9%; Spec 93.2% (DeLong P =0.004 vs clinical). Radscore (VOI1+VOI2) AUC = 0.962/0.926	Nomogram (visual)	Multiregional CTE radiomics (intestinal wall + largest LN) + clinical (involved segments) + endoscopy (longitudinal ulcer) via logistic nomogram; highest clinical benefit (DCA); Acc 89.5% superior to readers (66.7%-75.2%); good calibration; single-center Chinese cohort. Completes the 16 studies with visual nomogram for CD vs ITB differentiation
Lu et al[27], 2023	China	Retrospective; 1271 EHR patients (875 CD, 396 ITB)	CD vs ITB (colonoscopy text)	Free-text EHR (NLP)	Debiased TextCNN (distilled; integrated gradients)	Acc 0.83 (noisy value 0.70); superior to non-debiased	Integrated gradients; 39 phrases (17 guideline-based)	Trustworthy AI (accuracy + interpretability + robustness)
Chen et al[28], 2022	China	Retrospective; 287 cases (59 indicators)	ITB vs CD	Clinical + laboratory + endoscopic (multimodal)	Fusion correlation neural network	Average accuracy > 91%	Interpretable rules	Perianal fistulas, comb sign, cobblestone to CD; ascites to ITB
Weng et al[29], 2022	China	Retrospective; 200 patients (160 CD, 40 ITB)	ITB vs CD	Clinical + radiological (9 variables)	XGBoost (best); SHAP	AUC = 0.891; Sens 0.813; Spec 0.969; MCC 0.801	SHAP (global and local)	First SHAP use; perianal fistulas and comb sign point to CD
Zhu et al[30], 2021	China	Retrospective; 160 patients (93 CD, 67 ITB)	CD vs ITB (ileocecal)	CT enterography radiomics + clinical	Radiomics (9 GBDT features) + 2 clinical	AUC = 0.96 (train)/0.93 (value); superior to singles (P < 0.01)	Nomogram (visual)	Excellent bedside translation; helps treatment decisions
Kim et al[31], 2021	South Korea	Retrospective; 6617 colonoscopy images (211 CD, 299 BD, 217 ITB)	3-class CD vs BD vs ITB (pairwise)	Colonoscopy images (white-light)	DL CNN; Grad-CAM	3-class Acc 65.15% (all)/72.01% (typical); AUROC = 0.78-0.86	Grad-CAM	Best on typical images; helpful for inexperienced endoscopists
Tong et al[32], 2020	China	Retrospective; 6,399 patients (5,128 UC, 875 CD, 396 ITB)	3-class UC vs CD vs ITB (pairwise)	Free-text endoscopic reports (NLP)	RF and TextCNN	RF: UC-CD 0.89/0.84, UC-ITB 0.83/0.82, CD-ITB 0.72/0.77 (sens/spec)	None reported	Largest NLP cohort; excellent for inexperienced clinicians

CD: Crohn’s disease; GITB: Gastrointestinal tuberculosis; SMOTE: Synthetic minority over-sampling technique; ITB: Intestinal tuberculosis; WSI: Whole slide image; ATB: Abdominal tuberculosis; LN: Lymphadenitis; PTB: Pulmonary tuberculosis; EHR: Electronic health record; BD: Behçet’s disease; UC: Ulcerative colitis; CT: Computed tomography; MRE: Magnetic resonance elastography; ATR-FTIR: Attenuated total reflection-Fourier transform infrared; mid-IR: Middle infrared; CECT: Contrast-enhanced computed tomography; LN: Lymph node; VOI: Volume of interest; CI: Confidence interval; NLP: Natural language processing; AI: Artificial intelligence; ML: Machine learning; CNN: Convolutional neural network; DL: Deep learning; LASSO: Least absolute shrinkage and selection operator; XGBoost: Extreme gradient boosting; SHAP: SHapley additive exPlanations; LIME: Local interpretable model-agnostic explanations; LR: Logistic regression; GBDT: Gradient boosting decision tree; Grad-CAM: Gradient-weighted class activation mapping; RF: Random forest; TextCNN: Text convolutional neural network; Acc: Accuracy; AUROC: Area under the receiver operating characteristic curve; AUC: Area under curve; Sens: Sensitivity; Spec: Specificity; MCC: Matthews correlation coefficient; TB: Tuberculosis; AI: Artificial intelligence; MDT: Multidisciplinary team; GI: Gastrointestinal.

Table 2 Categorization of artificial intelligence and machine learning techniques according to research theme and representative studies in abdominal tuberculosis

Theme	Imaging/point data	Clinical/areal data	Key techniques	Representative original studies (n = 16)
Intestinal tuberculosis vs Crohn’s disease	Endoscopy images (white-light colonoscopy), CT/MRE radiomics, histopathology WSI, ATR-FTIR spectroscopy	Clinical/EHR parameters, nomograms, T-SPOT, pulmonary TB history	CNN/deep learning (U-Net, ResNet), radiomics (handcrafted + deep learning), XGBoost, few-shot learning (Xception), NLP/TextCNN (debiased), multimodal fusion, SHAP/LIME/attention maps explainability	Park et al[17], 2025; Cheng et al[18], 2024; Shu et al[19], 2024; Liu et al[21], 2024; Li et al[22], 2024; Lin et al[23], 2024; Gong et al[26], 2023; Lu et al[27], 2023; Chen et al[28], 2022; Weng et al[29], 2022; Zhu et al[30], 2021; Kim et al[31], 2021; Tong et al[32], 2020
Peritoneal tuberculosis vs carcinomatosis	CT (omental/peritoneal signs, ascites, lymph nodes, scalloping)	Omental/peritoneal imaging signs	Machine learning ensembles, radiomics (shape, first-order, texture features)	Pang et al[25], 2023
Nodal tuberculosis vs lymphoma	Contrast-enhanced CT (lymph node features)	Lymph-node morphological/textural features	Radiomics (logistic regression/SVM)	Shen et al[24], 2024
Multimodal integration	Fusion of MRE radiomics + colonoscopy images + histopathology WSI	Integrated clinical + endoscopic + radiological + pathological features	Multidisciplinary fusion models (LASSO + logistic regression), fusion correlation neural networks	Lu et al[20], 2024; Chen et al[28], 2022
Novel/emerging approaches	Few-shot endoscopic images, ATR-FTIR spectroscopy, noisy EHR text	Clinical parameters (for few-shot/explainable models)	Few-shot learning (dual transfer), spectroscopy + XGBoost/ML, debiased TextCNN (NLP), explainable AI (SHAP, LIME, integrated gradients, attention maps)	Lin et al[23], 2024; Li et al[22], 2024; Lu et al[27], 2023 (also embedded in multiple intestinal TB vs CD studies above)

CT: Computed tomography; MRE: Magnetic resonance elastography; WSI: Whole slide image; ATR-FTIR: Attenuated total reflection-Fourier transform infrared; EHR: Electronic health record; TB: Tuberculosis; CNN: Convolutional neural network; XGBoost: Extreme gradient boosting; NLP: Natural language processing; TextCNN: Text convolutional neural network; SHAP: SHapley additive exPlanations; LIME: Local interpretable model-agnostic explanations; SVM: Support vector machine; LASSO: Least absolute shrinkage and selection operator; ML: Machine learning; CD: Crohn’s disease.