💻 computer science

Toward Risk-Aware Clinical Decision Support in Gastrointestinal Endoscopy: A Tunable Asymmetric Loss with Uncertainty-Guided Referral

This study proposes a risk-aware deep learning framework for gastrointestinal endoscopy that utilizes a tunable asymmetric loss function, uncertainty-guided referral strategies, and lightweight architectures to achieve high-performance four-class risk stratification while prioritizing safety through the escalation of high-risk and pre-malignant cases.

Original authors: Azizur Rahman, Nakib Uddin Ahmed, Kadirur Rahman Chowdhury, Mehjabin Ferdous, J.S. Jannatun Nayema, Md. Shahadat Hossain, Akhlak Uz Zaman

Published 2026-10-06

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Azizur Rahman, Nakib Uddin Ahmed, Kadirur Rahman Chowdhury, Mehjabin Ferdous, J.S. Jannatun Nayema, Md. Shahadat Hossain, Akhlak Uz Zaman

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Inside the human body, the digestive tract is a long, winding tube that doctors examine using a flexible camera called an endoscope. This procedure, known as an endoscopy, allows physicians to see the lining of the stomach and intestines directly. It is a vital tool for finding problems early, such as inflammation or growths that could turn into cancer. However, the human eye can miss subtle signs, especially when a doctor is looking at hundreds of images in a single day or when a lesion is small and faint. To help, scientists have been developing computer programs that can look at these images and spot abnormalities. For years, these programs have mostly been trained to answer a simple yes-or-no question: is there a problem here, or is everything normal? While useful, this binary approach does not capture the full picture of patient care, where the urgency of a finding can vary wildly from a minor irritation to a life-threatening condition.

A new study from researchers at several universities seeks to move beyond this simple "yes or no" approach. Instead of just asking if a lesion exists, the team trained artificial intelligence to sort endoscopic images into four distinct levels of risk. They created categories for normal tissue, mild inflammation, pre-cancerous changes, and high-risk lesions that require immediate attention. To teach the computer to understand these differences, the researchers did not just show it pictures; they also taught it that some mistakes are far more dangerous than others. They built a system that penalizes the computer heavily if it misses a high-risk case, even if that means it occasionally flags a harmless image for extra review. This approach, which they call "risk-aware," aims to make the computer a safer partner for doctors, prioritizing patient safety over perfect accuracy.

The researchers tested their system using a large collection of real endoscopic images from a public database called HyperKvasir. They organized thousands of these images into their four risk groups based on guidelines from major medical societies. To see which type of computer brain worked best, they compared three different lightweight models. These models are designed to be efficient, using fewer than eight million adjustable settings, which makes them faster and less demanding on computer hardware than massive systems. One model learned by looking at small patches of an image, another by connecting layers of features, and a third by analyzing the whole picture at once. The team trained these models to recognize the four risk levels, using a special scoring method that made the computer pay extra attention to the most dangerous categories.

When the team tested the results, they found that the system could successfully distinguish between the different risk levels. The best-performing model correctly identified the risk category in about 84 percent of the test images. More importantly, the system was exceptionally good at catching the most dangerous cases. In their tests on the internal dataset, the model did not predict any ground-truth high-risk images as normal or low-risk; every single high-risk case was either correctly identified or flagged for a doctor to review. This safety net was achieved by a specific rule: if the computer thought an image was pre-cancerous or high-risk, it automatically sent it to a human doctor for a second look, regardless of how confident the computer felt about its own answer. This rule meant that about 43 percent of the images in their test set could be cleared automatically as safe, while the rest were escalated for human review.

The study also looked at how well these computer models would work on images they had never seen before, taken from a different hospital with different equipment. When tested on this new set of images, the models' overall accuracy dropped, which is a common challenge for medical AI. However, even with this drop in performance, the system remained very good at spotting the high-risk cases, maintaining an AUC of over 0.97 for the most dangerous tier. This suggests that while the models might struggle with the finer details of different types of tissue, they learned to recognize the specific visual signs of the most critical conditions well enough to transfer that knowledge to new environments. The researchers also used a technique to show which parts of the image the computer was focusing on. The heat maps revealed that the models were looking at the actual tissue patterns and lesions, rather than random parts of the screen, which gives doctors more confidence in the system's decisions.

Despite these promising results, the authors are careful to state that this is a simulation, not a finished medical product. The study was conducted entirely on computers using existing data, without any real patients or doctors interacting with the system in a live setting. The researchers emphasize that their four-level risk system is a simplified version of real-world medicine, designed to test a specific idea rather than to replace a doctor's diagnosis. They found that while the system was excellent at catching the most dangerous cases in their specific test set, this result is specific to that dataset and does not guarantee the same performance in clinical practice. They also found that the system sometimes struggled with images showing mild inflammation, which can look very similar to normal tissue. This difficulty highlights that the computer still needs human guidance, particularly for the middle ground between healthy and dangerous.

The work demonstrates that it is possible to build an artificial intelligence system that understands the difference between a minor issue and a medical emergency. By teaching the computer to value safety over simple accuracy, the researchers created a tool that could potentially reduce the workload for doctors by automatically clearing safe cases while ensuring that every potential danger gets a human second look. The study concludes that while the technology shows great promise, it requires further testing in real hospitals with diverse patients and equipment before it can be trusted to assist in actual medical procedures. The path forward involves refining these models, testing them across different hospitals, and ensuring they can handle the complex and varied reality of human biology before they become a standard part of medical care.

Technical Summary: Toward Risk-Aware Clinical Decision Support in Gastrointestinal Endoscopy

Problem Statement
Current computer-aided detection (CAD) systems for gastrointestinal (GI) endoscopy predominantly rely on binary classification (lesion vs. no lesion) or simple multi-class detection. This binary approach fails to capture the nuanced clinical risk associated with different endoscopic findings, which dictates distinct management pathways (e.g., surveillance, biopsy, resection) as outlined by guidelines from the American College of Gastroenterology (ACG) and the European Society of Gastrointestinal Endoscopy (ESGE). Furthermore, existing deep learning models often treat all classification errors as equivalent, ignoring the asymmetric clinical consequences where missing a high-risk lesion (false negative) is far more detrimental than a false alarm. There is a need for a unified framework that integrates risk-oriented stratification, cost-sensitive learning, uncertainty estimation, and explainability within a computationally efficient architecture suitable for clinical decision support.

Methodology
The study proposes a four-class risk stratification framework applied to GI endoscopic images, categorizing findings into: Normal, Inflammatory, Pre-malignant, and High-Risk. These categories were derived by mapping the 23 classes of the HyperKvasir dataset to risk tiers based on ACG/ESGE management pathways, excluding non-pathological classes (e.g., bowel preparation quality).

Asymmetric Endoscopy Loss (AEL): The core methodological contribution is the introduction of AEL, a tunable, cost-sensitive weighted cross-entropy loss function. Unlike standard cross-entropy or Focal Loss, AEL explicitly encodes a risk hierarchy into the training objective. It assigns a fivefold penalty ( $w=5.0$ ) to misclassifications of the High-Risk tier, a 3.5x penalty for Inflammatory, and a 3.0x penalty for Pre-malignant, relative to a baseline of 1.0 for Normal. This allows for a trade-off between overall accuracy and safety-oriented recall for high-risk cases.
Architectures: Three lightweight architectures (<8 million parameters) were benchmarked under identical conditions to compare Convolutional Neural Networks (CNNs) and Transformers: DenseNet-121, EfficientNet-B0, and DeiT-Tiny. All models were pre-trained on ImageNet-1k and fine-tuned.
Uncertainty and Referral: Predictive uncertainty was estimated using Monte Carlo (MC) Dropout with 30 stochastic forward passes during inference. A rule-based referral strategy was simulated: any prediction classified as Pre-malignant or High-Risk was automatically escalated for endoscopist review, regardless of the model's confidence score. Additionally, predictions with confidence below a threshold ( $\tau = 0.75$ ) were also referred.
Explainability: Gradient-weighted Class Activation Mapping (GradCAM) was employed to visualize the image regions driving model predictions.
Evaluation Protocol: Models were trained on the HyperKvasir dataset (7,633 images) and evaluated on a held-out internal test set (1,145 images). Zero-shot generalization was tested on the independent Kvasir-v2 dataset. Performance was assessed via macro F1-score, per-class AUC-ROC, and High-Risk recall. Robustness was further analyzed using five-fold cross-validation and McNemar's statistical tests for paired comparisons.

Key Results

Classification Performance: On the internal HyperKvasir test set, DenseNet-121 achieved the highest macro F1-score of 0.8419 with a mean AUC-ROC of 0.9730. EfficientNet-B0 followed closely (0.8390), while DeiT-Tiny scored 0.8106. McNemar's tests indicated no statistically significant difference between DenseNet-121 and EfficientNet-B0 ( $p=0.911$ ), though EfficientNet-B0 significantly outperformed DeiT-Tiny ( $p=0.034$ ).
High-Risk Safety: All architectures demonstrated high recall for the High-Risk category (ranging from 0.9505 to 0.9659 on the internal set). Crucially, under the proposed referral rule, zero ground-truth High-Risk images were predicted as Normal or Inflammatory in the internal test set, meaning no High-Risk cases were "auto-cleared."
Workload Simulation: At a confidence threshold of $\tau=0.75$ , 43.1% of test images were eligible for automatic clearance (i.e., predicted as Normal or Inflammatory with high confidence), while the remaining 56.9% were flagged for review.
Generalization: Performance dropped on the external Kvasir-v2 dataset (macro F1 drops of 0.04–0.10), confirming the presence of domain shift. However, High-Risk recall remained robust (>0.97) across all models on the external dataset.
Ablation and Sensitivity: An ablation study comparing AEL against standard Cross-Entropy and Focal Loss (using random initialization and no CLAHE) showed that while AEL did not yield the highest macro F1-score (0.7821 vs. 0.8361 for Focal Loss), it achieved the highest High-Risk recall (0.9587). This confirms AEL's role as a mechanism to prioritize safety over aggregate accuracy.
Explainability: GradCAM visualizations showed that CNNs produced sharper, localized attention on lesions, while DeiT-Tiny captured broader patch-level patterns. Attention maps generally aligned with mucosal textures and lesion boundaries.

Significance and Claims
The authors position this work as a computational investigation rather than a clinically validated diagnostic system. The study's significance lies in demonstrating the feasibility of integrating a risk-oriented, guideline-informed four-class schema with asymmetric cost-sensitive learning and uncertainty-aware referral within a lightweight deep learning framework.

Key claims include:

Risk-Aware Formulation: Moving beyond binary detection to a four-tier risk stratification provides a more clinically informative representation of endoscopic findings.
Tunable Safety: The AEL function offers a controllable mechanism to trade off overall accuracy for safety-oriented recall, allowing the system to be tuned for specific clinical priorities without requiring complex post-hoc thresholding.
Referral Strategy: Combining uncertainty estimation with a rule-based escalation for high-risk categories can effectively simulate a reduction in endoscopist workload (via auto-clearance) while maintaining a safety net that prevents high-risk cases from being missed in the simulated environment.
Architectural Comparison: The study provides a controlled comparison of CNN and Transformer architectures under a strict parameter budget, showing that lightweight models can achieve competitive performance in this specific risk-stratification task.

The paper explicitly acknowledges limitations, including the study-defined nature of the risk schema (not a replacement for histopathology), the lack of patient-level data in the datasets (risk of data leakage), and the retrospective nature of the workload simulation. The authors conclude that while the framework shows promise, prospective multi-institutional validation and improved probability calibration are necessary before clinical deployment.

More like this