📄 medicine

Beyond Binary: Four-Class Risk Stratification from Gastrointestinal Endoscopy Using Asymmetric-Cost Lightweight CNN–Transformer Learning

This paper introduces a lightweight CNN–Transformer benchmark that utilizes an asymmetric-cost loss function and Monte Carlo Dropout to achieve four-class gastrointestinal lesion risk stratification aligned with clinical guidelines, successfully eliminating missed high-risk cases while enabling automated clearance of nearly half of the cases without endoscopist review.

Original authors: Azizur Rahman, Nakib Uddin Ahmed, Kadirur Rahman Chowdhury, Mehjabin Ferdous, Md. Shahadat Hossain, Md. Amir Hossain

Published 2026-07-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Azizur Rahman, Nakib Uddin Ahmed, Kadirur Rahman Chowdhury, Mehjabin Ferdous, Md. Shahadat Hossain, Md. Amir Hossain

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

Imagine a world where doctors use a tiny camera on a long, flexible tube to peek inside our stomachs and intestines. This is called an endoscopy, and it's the best way to spot trouble before it turns into something serious, like cancer. For a long time, the computers helping doctors with this task were a bit like a simple light switch: they could only tell you "yes, there's a problem" or "no, everything looks fine." But real life isn't that black and white. Sometimes a spot is just a little red and angry (inflammation), sometimes it's a warning sign that might turn bad later (pre-malignant), and sometimes it's a ticking time bomb that needs immediate removal (high-risk). Treating all these different situations the same way is like using a sledgehammer to crack a nut; it misses the nuance doctors need to make the right choice. This paper dives into the world of medical artificial intelligence (AI) to see if we can build a smarter helper that doesn't just shout "danger," but actually sorts patients into the right level of care, just like a seasoned doctor would.

The researchers behind this study wanted to upgrade the AI from a simple "yes/no" switch to a four-level traffic light system. They built a new kind of digital brain that looks at endoscopy images and sorts them into four buckets: Normal (keep an eye on it later), Inflammatory (needs medicine), Pre-malignant (needs a biopsy to check for early cancer), and High-Risk (needs immediate action). To do this, they taught three different types of lightweight AI models—think of them as compact, efficient digital brains that don't need a supercomputer to run—using a special set of rules called the "Asymmetric Endoscopy Loss." This is a fancy way of saying the computer was taught that making a mistake on a "High-Risk" patient is five times worse than making a mistake on a "Normal" patient. It's like a security guard who is very lenient about people just walking by but goes into full panic mode if they see someone trying to break a window.

The team tested their system on thousands of real endoscopy images. They found that their lightweight models were surprisingly good at the job. The best performer, a model called DenseNet-121, correctly sorted about 84% of the cases overall. But the real magic happened when they added a safety net called "Monte Carlo Dropout." Imagine asking the AI to look at the same picture 30 times and asking, "Are you sure?" If the AI hesitates or if it thinks a spot is dangerous, the system automatically flags it for a human doctor to double-check. Using this method, the system successfully caught every single High-Risk lesion in their test group—zero were missed. At the same time, it was able to automatically clear about 44.9% of the cases (the ones that were definitely normal or just slightly inflamed) without needing a doctor to look at them at all. This means doctors could save nearly half their time on routine checks while feeling confident that the dangerous cases were never overlooked.

The study also showed that these smart models could learn from one set of images and still do a decent job on a completely different set of images from another hospital, though they weren't perfect at it yet. The researchers used a tool called GradCAM to peek inside the AI's "mind," and they found that the computer was actually looking at the right things—like the texture of the tissue and the shape of the lining—rather than getting distracted by weird shadows or text on the screen. While the AI still struggled a bit with the "Inflammatory" category (often confusing it with normal tissue), the researchers noted that this wasn't a safety failure; it just meant a doctor might check a harmless spot a bit more closely than necessary, which is safer than missing a dangerous one.

In short, this paper suggests that we don't need massive, heavy computers to build a smart endoscopy assistant. By using smaller, efficient models and teaching them to care more about missing a dangerous cancer than about a false alarm, we can create a system that speeds up doctor's work without sacrificing safety. It's a promising step toward a future where AI acts as a tireless, ultra-cautious partner, ensuring that the most critical patients get the attention they need immediately, while letting doctors focus their energy where it matters most.

Technical Summary: Beyond Binary: Four-Class Risk Stratification from Gastrointestinal Endoscopy Using Asymmetric-Cost Lightweight CNN–Transformer Learning

Problem Statement
Current AI-assisted endoscopy systems predominantly operate on a binary "lesion-present/absent" framework. The authors argue this design is fundamentally misaligned with published clinical practice guidelines (specifically those from the American College of Gastroenterology and the European Society of Gastrointestinal Endoscopy), which categorize gastrointestinal (GI) findings into tiered risk levels requiring distinct clinical actions. Furthermore, binary systems treat all misclassification errors equally, ignoring the critical clinical asymmetry where missing a high-risk lesion (potentially leading to invasive cancer) has far more severe consequences than over-triaging a normal finding. Existing deep learning approaches also often rely on heavy models unsuitable for clinical hardware or fail to address the specific cost asymmetry of GI risk stratification.

Methodology
The study proposes a lightweight, four-class risk stratification framework designed to run on standard clinical hardware. The methodology encompasses the following components:

Clinical Risk Schema: The authors mapped 21 of the 23 finding classes from the HyperKvasir dataset into four actionable tiers: Normal (routine surveillance), Inflammatory (medical management), Pre-malignant (biopsy/intensified surveillance), and High-Risk (urgent intervention/resection). Two non-pathological classes were excluded.
Model Architectures: Three lightweight architectures, each with fewer than 8 million parameters, were benchmarked:
- DenseNet-121 (CNN)
- EfficientNet-B0 (CNN)
- DeiT-Tiny (Vision Transformer)
  All models were pre-trained on ImageNet-1k and fine-tuned with task-specific classification heads.
Asymmetric Endoscopy Loss (AEL): A novel weighted cross-entropy loss function was introduced to encode clinical risk asymmetry directly into the training objective. The weight vector $w = [1.0, 3.5, 3.0, 5.0]$ assigns a 5× penalty to the misclassification of High-Risk lesions compared to Normal findings. The weight for Inflammatory lesions (3.5) was also elevated to partially correct for severe class imbalance in the dataset.
Preprocessing and Training: Images underwent Contrast-Limited Adaptive Histogram Equalization (CLAHE) to enhance mucosal contrast. Training utilized a WeightedRandomSampler to balance batches.
Uncertainty Quantification and Referral Protocol: Monte Carlo (MC) Dropout was employed during inference (30 stochastic passes) to estimate predictive uncertainty. A safety-oriented referral protocol was established:
1. Any prediction classified as Pre-malignant or High-Risk is automatically escalated for endoscopist review, regardless of confidence.
2. Predictions with confidence below a threshold ( $\tau = 0.75$ ) are also escalated.
3. Only high-confidence Normal or Inflammatory predictions are auto-cleared.
Explainability: GradCAM was used to generate saliency maps, verifying that model attention aligns with clinically relevant mucosal features.

Key Results

Classification Performance: On the HyperKvasir test set, DenseNet-121 achieved the highest macro F1-score (0.8371) and mean AUC-ROC (0.9704). EfficientNet-B0 demonstrated superior zero-shot generalization on the independent Kvasir-v2 dataset (macro F1 = 0.7756) with the smallest performance drop compared to in-domain results.
High-Risk Detection: All architectures achieved High-Risk recall above 94% under direct classification. Crucially, under the MC Dropout referral protocol, zero High-Risk lesions were missed on the internal test set, as the mandatory escalation rule captured all Pre-malignant and High-Risk predictions.
Workload Reduction: The system successfully auto-cleared 44.9% of cases (high-confidence Normal and Inflammatory findings), potentially reducing endoscopist review burden by approximately 45% without compromising safety.
Ablation and Sensitivity: An ablation study confirmed that the AEL loss function effectively prioritizes safety-critical classes. While Focal Loss achieved a slightly higher macro F1 in a controlled protocol, AEL maintained competitive High-Risk recall (0.9546) while enforcing the specific clinical cost asymmetry. Sensitivity analysis identified the 5.0 penalty weight as the optimal operating point.
Explainability: GradCAM heatmaps confirmed that models focused on clinically relevant features (e.g., mucosal irregularities, dye boundaries) rather than artifacts. CNNs produced sharper, localized attention, while the Transformer showed broader, context-aware patterns.

Significance and Claims
The paper positions this work as a computational benchmark establishing the feasibility of guideline-aligned, lightweight AI for GI endoscopy, rather than a validated clinical deployment. Its primary contributions are:

Reframing the Task: Moving from binary lesion detection to a four-class risk stratification task that directly mirrors ACG/ESGE clinical guidelines.
Asymmetric Cost Learning: Introducing the AEL loss function to explicitly encode the survival-cost asymmetry of missing High-Risk lesions, a feature absent in prior GI classification work.
Safety-First Architecture: Demonstrating that lightweight models (<8M parameters) can achieve high performance and that integrating MC Dropout with a hard safety constraint (mandatory escalation of high-risk tiers) can structurally guarantee zero missed High-Risk lesions.
Hardware Feasibility: Proving that both Convolutional and Transformer architectures can operate within the parameter constraints of standard clinical endoscopy hardware while maintaining robust performance.

The authors acknowledge limitations, including the single-institution source of the primary training data (HyperKvasir), the simplification of complex clinical nuances into four tiers, and the need for prospective multi-site validation before clinical deployment. They note that while the models are lightweight, real-time inference latency on specific clinical hardware (e.g., embedded GPUs) requires further characterization.

Technical Summary: Beyond Binary: Four-Class Risk Stratification from Gastrointestinal Endoscopy Using Asymmetric-Cost Lightweight CNN–Transformer Learning

More like this