🤖 Generated with AI assistance · Does not replace individualized medical advice.
Artificial intelligence promises to revolutionize the histopathological diagnosis of Hirschsprung disease, but a systematic review reveals that 69% of current studies have a high risk of bias, limiting their immediate clinical application.
Keyfacts
| Studies reviewed | AI performance | Time reduction | Risk of bias |
|---|---|---|---|
| 13 studies (2016-2025) PMID: 42667407 | >90% in ganglion cell detection PMID: 42667407 | 50-95% PMID: 42667407 | 69% high risk PMID: 42667407 |
Context / Why it matters
Hirschsprung disease is characterized by the absence of ganglion cells in the distal large intestine, causing severe constipation and requiring accurate histopathological diagnosis. The conventional method, based on microscopic evaluation of rectal biopsies, is slow, subjective, and dependent on the pathologist’s experience. In this scenario, artificial intelligence (AI) emerges as a promising tool to increase accuracy, reduce interobserver variability, and accelerate clinical decision-making. However, its integration into real-world practice requires a rigorous assessment of methodological quality and risk of bias. This systematic review, following PRISMA 2020 guidelines, analyzes the current state of AI in Hirschsprung diagnosis and highlights the gaps that must be closed before its clinical adoption.
Details
The review identified 13 studies that evaluated machine learning and deep learning techniques for the analysis of histopathological images of Hirschsprung disease. The approaches were grouped into three main categories, each with distinct characteristics and performance:
| Approach | Number of studies | Performance (ganglion cell detection) | Diagnostic time reduction |
|---|---|---|---|
| Traditional image processing | 4 PMID: 42667407 | Below 90% PMID: 42667407 | Not quantified PMID: 42667407 |
| Convolutional neural networks (CNN) | 6 PMID: 42667407 | >90% PMID: 42667407 | 50-95% PMID: 42667407 |
| Transformer models | 3 PMID: 42667407 | >90% PMID: 42667407 | 50-95% PMID: 42667407 |
Deep learning models (CNN and Transformer) consistently outperformed traditional methods, achieving accuracy above 90% in ganglion cell detection and reducing diagnostic time by 50% to 95%. This improvement is attributed to the networks’ ability to extract complex features from images without the need for extensive manual preprocessing.
Limitations
⚠️ Key limitations
69% of studies (9 of 13) present a high risk of bias according to the QUADAS-AI and PROBAST frameworks. The main sources of bias are: small sample sizes, patch-level data partitioning (which can cause information leakage between training and test sets) and the absence of external test sets. These factors raise serious concerns about overfitting and generalization of the models, limiting their real clinical utility.
Clinical takeaway
AI can support the diagnosis of Hirschsprung disease, but external validation and multicenter datasets are required before its routine clinical implementation.
Full reference
Wan JH, Hasikin K, Mun KS, Tan YW. Advancing hirschsprung disease diagnosis: a systematic review of the development and application of artificial intelligence in histopathological analysis. Pediatr Surg Int. 2026; PMID: 42667407, DOI: 10.1007/s00383-026-06576-3.