💻 computer science

CogAlign-CausalX: mask-grounded polyp analysis with closed-form explanation-alignment scoring

CogAlign-CausalX is a polyp analysis pipeline that treats segmentation masks as intrinsic explanations for morphological classification, validated by a fast, closed-form alignment scoring metric (XAlign) that achieves competitive performance on the Kvasir-SEG dataset while acknowledging limitations in absolute segmentation accuracy and the need to pair alignment scores with traditional overlap measures.

Original authors: G. Sudha Gowd, S. Vasundra

Published 2026-10-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: G. Sudha Gowd, S. Vasundra

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

Every year, millions of people face the risk of colorectal cancer, a disease that often begins as a small, harmless-looking growth called a polyp inside the colon. Doctors use a procedure called a colonoscopy to look inside the colon and remove these growths before they can turn into something dangerous. For years, researchers have been teaching computers to help spot these polyps in the video images taken during the procedure. These computer programs have become very good at drawing a line around a polyp, separating it from the rest of the tissue. However, a major problem remains: while the computer can point to the polyp, it cannot easily explain why it thinks that is a polyp. In high-stakes medicine, a doctor needs to trust the machine's reasoning, not just its final guess. If a computer says "this is a polyp," the doctor needs to know if the machine is looking at the right features or if it is just guessing based on a lucky pattern.

A team of researchers has proposed a new way to solve this trust problem. Instead of building a complex system that guesses first and then tries to explain itself later, they designed a computer pipeline where the explanation is built into the very first step of the thinking process. Their method, which they call CogAlign-CausalX, works by forcing the computer to make its decision using only the specific area it has already identified. Imagine a doctor looking at a photo of a colon and covering everything except the suspicious spot with a piece of paper; the doctor then makes a judgment based solely on what is visible under that paper. In this system, the computer does the same thing. It first draws a mask, or a digital outline, around a potential polyp. Then, it throws away all the rest of the image and looks only at the pixels inside that outline to decide what kind of growth it is. Because the computer cannot see anything outside that outline, the outline itself becomes the explanation. If the computer gets the answer right, it is because it correctly identified the right area to look at.

To make sure this approach actually works, the researchers needed a way to measure how good the computer's outline was, not just by looking at it, but by giving it a score. They developed a new scoring method that checks three specific things about the outline. First, it checks if the computer's outline stays mostly inside the actual polyp. Second, it checks how closely the shape of the computer's outline matches the shape drawn by a human expert. Third, it checks if the computer's outline spills over into the healthy tissue surrounding the polyp. By combining these three checks into a single number, the researchers can tell, in a fraction of a second, how well the computer's "reasoning" aligns with human expertise. This is much faster than older methods that try to guess the computer's confidence by running the same image through the system dozens of times with slight random changes, a process that takes minutes instead of milliseconds.

The researchers tested their system on a collection of 1,000 real colon images that had been carefully checked and labeled by medical experts. They split these images into groups to train the computer and then to test it, making sure the groups did not overlap. The results showed that the system was quite effective. It successfully identified the polyps in most cases, with an accuracy score that placed it above several older computer models, though it was still slightly behind the very newest and most powerful systems available. More importantly, the system provided a clear, numerical score for its own explanations. The researchers found that when the computer made a mistake, it was usually a small one, such as guessing the polyp was slightly larger or smaller than it really was, rather than a complete misunderstanding of what it was looking at.

However, the researchers were careful to point out a specific weakness in their new scoring method. They discovered that the score could be tricked by a very cautious computer. If the computer drew a small outline that stayed perfectly inside a large polyp but missed most of the polyp's actual size, the score would still be very high. This is because the score rewards the computer for staying inside the safe zone and not spilling over into healthy tissue, but it does not heavily punish the computer for failing to cover the whole polyp. The researchers showed that a computer could miss two-thirds of a polyp and still receive a high score if it was just very careful about not touching the healthy tissue. This finding is crucial because it means the score cannot be used alone to judge the system. It must always be looked at alongside a standard measure of how much of the polyp was actually found.

The study also highlighted the limits of what this specific test could tell us. The computer was trained and tested on a single set of images, and the researchers used only one specific setup for their experiments. They noted that the system was not fast enough to run in real-time during a live medical procedure on the computers they used, meaning it is currently better suited for reviewing images after the fact rather than helping a doctor in the moment. Furthermore, the images they used were single, clear snapshots, not the moving, sometimes blurry video that doctors see during an actual procedure. The researchers emphasized that while their method offers a new way to measure trust in computer vision, it is not a finished product ready for the clinic. It is a proof of concept that shows how to build a system where the explanation is part of the decision, and it provides a clear roadmap for what needs to be fixed before such a system could be trusted with real patients.

In the end, this work is less about making the computer smarter at finding polyps and more about making the computer's thinking process visible and measurable. By forcing the computer to rely only on the area it highlights, the researchers created a system where the "why" is as important as the "what." They proved that you can build a machine that explains itself by design, rather than by adding a separate explanation step later. While their specific scoring tool has a blind spot regarding how much of a polyp is missed, the core idea—that the explanation should be the input, not the output—offers a promising path forward. The researchers have made their code and data available so that others can check their work, verify the numbers, and build upon their findings, ensuring that the next generation of medical AI is not just accurate, but also understandable.

Technical Summary: CogAlign-CausalX

Problem Statement
Current automated polyp segmentation systems achieve high overlap metrics (Dice coefficients > 0.90) on benchmarks like Kvasir-SEG, yet their explanations remain largely qualitative. Standard pipelines produce a segmentation mask and a classification label independently, with explanations (e.g., class activation maps) supplied post-hoc. These post-hoc explanations are often assessed only by visual inspection of favorable examples, a practice criticized for lacking fidelity; saliency maps can survive randomization of model weights, and visual plausibility does not equate to measurement. The core difficulty lies in the lack of a constrained relationship between the segmentation and the classification, and the absence of a quantitative metric to evaluate the quality of the explanation itself.

Methodology: CogAlign-CausalX Pipeline
The authors propose a pipeline where the explanation is a structural property of the computation rather than a post-hoc attribution. The system operates in four stages:

Stage 1 (Segmentation): An EfficientNet-B0 encoder with a U-Net decoder generates a probability map. The network is trained using a composite loss function combining binary cross-entropy, soft Dice loss, and a contour term penalizing disagreement between the Laplacian responses of the prediction and the target.
Stage 1.5 (Deterministic Post-processing): The probability map is binarized at a threshold ( $\tau^* = 0.55$ ) selected exclusively on the validation split. A $3 \times 3$ average pooling step removes isolated pixels, ensuring the process is fully deterministic and reproducible.
Stage 2 (Region-of-Interest Extraction): The binary mask gates the input image via a Hadamard product. Crucially, the mask is detached from the autograd graph before this operation. This ensures the classifier receives only the pixels within the predicted mask; pixels outside are identically zero. Thus, the classifier's reasoning is structurally constrained to the region selected by the segmenter.
Stage 3 (Morphological Classification): A lightweight ResNet-18 head classifies the masked region into three morphological tiers (focal, intermediate, extensive) based on relative lesion area.
Stage 4 (Explanation Alignment Scoring): The paper introduces XAlign, a closed-form, parameter-free metric to score the binary mask against expert annotation. It combines three terms:
- Regional Concentration (RC): The fraction of predicted mass falling inside the lesion.
- Boundary Adherence (BA): A measure of area agreement between prediction and annotation, normalized by image size.
- Dispersion Penalty (DP): An exponential penalty for predicted mass spilling outside the lesion.
  The final score is the unweighted mean of these three terms.

Key Contributions

Structural Explanation: The pipeline eliminates post-hoc attribution by making the segmentation mask the classifier's sole visual input. The mask is the explanation.
Closed-Form Scoring: XAlign provides a computable alignment score (0.42 ms per image) that is orders of magnitude faster than sampling-based uncertainty estimation (e.g., Monte-Carlo dropout), which requires ~16.4 seconds for 30 passes.
Rigorous Validation: The study employs a stratified 700/150/150 partition of Kvasir-SEG with MD5-verified disjointness, validation-only threshold selection, and bias-corrected bootstrap confidence intervals for all metrics.
Blind Spot Disclosure: The authors explicitly characterize the metric's failure modes, specifically its inability to penalize false negatives (under-segmentation) when the prediction lies entirely within the lesion.

Results
On the held-out test set (150 frames):

Segmentation: Achieved a Dice coefficient of 0.8472 (95% CI: 0.8191–0.8744) and an IoU of 0.7652.
Classification: Tier accuracy was 0.8067. All 29 misclassifications fell into adjacent tiers, indicating errors were geometric rather than catastrophic.
Alignment: The mean XAlign score was 0.8404. The score correlates with Dice ( $r=0.830$ ) but captures distinct variance.
Failure Analysis: Controlled experiments revealed a critical blind spot: a prediction occupying half the lesion area but lying entirely inside it received an XAlign score of 0.9821 (Dice = 0.667). This occurs because the metric lacks a false-negative term normalized by lesion mass.
Ablation: Removing the Laplacian boundary term reduced Dice by 0.0364 but increased the alignment score, confirming that the metric rewards the conservative, tighter masks produced without boundary supervision.

Significance and Claims
The paper positions its contribution not as a state-of-the-art segmentation model, but as a framework for measurable explainability. The authors acknowledge that their segmentation accuracy (Dice ~0.85) lags behind current transformer and state-space baselines (Dice ~0.91–0.92) due to the use of a smaller backbone (EfficientNet-B0) and limited training data.

The primary significance lies in the shift from qualitative visual inspection to quantitative evaluation of explanations. The authors argue that for high-stakes medical decisions, explanations should be constrained by construction (via gating) rather than bolted on. However, they maintain a modest stance regarding the XAlign metric: it is a component of an evaluation, not a standalone replacement for overlap measures. The paper explicitly states that XAlign must be reported alongside an overlap metric (like Dice) because it cannot detect conservative under-segmentation. The work concludes that future improvements require adding a lesion coverage term to the score, retraining with higher-capacity encoders, and validating on external cohorts with patient-level separation.

More like this