🤖 AI

Dissecting Model Failures in Abdominal Aortic Aneurysm Segmentation through Explainability-Driven Analysis

This paper proposes an Explainable AI (XAI) guided encoder shaping framework that leverages attribution-based focus maps to align model attention with predictions and refine inference, thereby significantly improving the segmentation of complex abdominal aortic aneurysms by mitigating failures caused by irrelevant structures and low-contrast targets.

Original authors: Abu Noman Md Sakib, Merjulah Roby, Zijie Zhang, Satish Muluk, Mark K. Eskandari, Ender A. Finol

Published 2026-03-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Abu Noman Md Sakib, Merjulah Roby, Zijie Zhang, Satish Muluk, Mark K. Eskandari, Ender A. Finol

Original paper licensed under CC BY 4.0 (http://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 you are trying to find a very specific, thin, and faintly glowing thread inside a tangled ball of yarn that also contains other similar-looking threads, knots, and shadows. This is essentially what a computer program is trying to do when it looks at a CT scan of a human abdomen to find an Abdominal Aortic Aneurysm (AAA).

An AAA is a dangerous bulge in the main artery. Doctors need to trace the exact outline of this bulge (the "outer wall") and the hollow space inside it (the "lumen") to decide if surgery is needed. But these walls are often paper-thin and look very similar to the surrounding tissue.

Here is how this paper solves the problem, using simple analogies:

The Problem: The "Distracted Student"

Current AI models (like U-Net or the famous "Segment Anything Model") are like brilliant but easily distracted students.

The Issue: When looking at a complex scan, these models often get confused. They might focus on a nearby vein instead of the aneurysm, or they might "leak" their attention into the background, drawing a messy, blobby shape instead of a precise outline.
The Old Way: Traditionally, we taught these models by showing them the right answer and saying, "You were wrong here, fix it." But the models didn't understand why they were wrong. They just tweaked their math to get the score up, without actually learning to look at the right thing. They were memorizing the test answers without understanding the lesson.

The Solution: The "Teacher with a Highlighter"

The authors propose a new method called XAI-SAM. Think of this as giving the student a high-tech highlighter and a strict teacher who watches where the student is looking, not just what they write.

The system works in three main steps:

1. The "Focus Map" (The Highlighter)

The researchers added a special tool that creates a "Focus Map" (called an XAI field).

Analogy: Imagine the AI is reading a book. Usually, we only check if it got the summary right. With this new tool, we can see exactly which words the AI is staring at while it reads.
The Fix: If the AI is staring at a nearby vein (the wrong word) but trying to outline the aneurysm, the system sees this mismatch immediately. It forces the AI to align its "gaze" with its "answer." If it says "This is the aneurysm," its eyes must be looking only at the aneurysm.

2. The "Confidence Filter" (The Bouncer)

The system also adds a "bouncer" at the exit of the AI's brain.

Analogy: Before the AI draws the final picture, it runs the image through a filter that asks, "Are you sure about this part?"
The Fix: If the AI is unsure or sees a confusing shadow, the bouncer suppresses that part of the image. It stops the AI from drawing lines in places where it's just guessing, effectively cutting out the "noise" and keeping only the confident, clear structures.

3. The "Consistency Check" (The Flip-Book)

Medical scans are like a stack of thousands of thin slices (like pages in a book).

Analogy: If you flip through a book, the drawing of a person shouldn't suddenly jump from the left side of the page to the right side between two pages. That would look like a glitch.
The Fix: The system was trained to look at pairs of slices and ask, "Does this shape make sense next to the one before it?" If the AI draws a shape that suddenly jumps or disappears in a way that doesn't make anatomical sense, the system penalizes it. This ensures the 3D shape is smooth and logical, not a jagged mess.

The Results: From "Good Enough" to "Expert"

When they tested this new method on difficult cases where other models failed:

Old Models: Often missed the thin walls, drew the wrong shape, or got confused by nearby blood vessels.
XAI-SAM: It acted like a master surgeon. It found the thin, hard-to-see walls with incredible precision. It didn't get distracted by nearby veins, and the 3D shape it built was smooth and consistent.

Why This Matters

This paper is a big deal because it changes how we train AI. Instead of just saying "Get the right answer," we are now teaching the AI how to look.

By forcing the AI to pay attention to the right things (the "Focus Map") and checking that its logic makes sense from slice to slice, we get a system that is not only more accurate but also more trustworthy. It's the difference between a student who guesses the right answer by luck and a student who truly understands the material.

In short: They taught the AI to stop daydreaming, focus on the right target, and double-check its work, resulting in a much safer and more reliable tool for doctors to use.

1. Problem Statement

Context: Abdominal Aortic Aneurysm (AAA) segmentation from Computed Tomography Angiography (CTA) images is critical for surgical planning and risk assessment. The task involves segmenting two specific structures: the outer aneurysm wall and the lumen.
The Challenge: While modern deep learning models (e.g., U-Net variants, MedSAM, SAM) perform well on clean data, they frequently fail in complex clinical scenarios characterized by:

Thin, low-contrast targets: The aneurysm wall often blends with surrounding soft tissue.
Distractors: Models often misallocate probability mass to nearby vessels, calcifications, or irrelevant structures.
Inconsistency: Predictions often exhibit abrupt, anatomically implausible jumps between consecutive slices.
Root Cause: The authors hypothesize that these failures stem from a misalignment between the model's internal focus (attribution) and the actual target structures. Standard loss functions (e.g., Dice, Cross-Entropy) optimize pixel-wise overlap but do not explicitly constrain where the encoder looks during the decision-making process.

2. Methodology: XAI-SAM Framework

The authors propose XAI-SAM, an Explainable AI (XAI)-guided framework built upon a Segment Anything Model (SAM) backbone. The core innovation is treating the encoder's internal attribution map as a primary training signal rather than a post-hoc diagnostic tool.

A. Architecture Overview

Backbone: Utilizes a SAM-style encoder-decoder with a Vision Transformer (ViT) encoder.
XAI Field Generation: A gradient-based attribution map (the "XAI field") is computed from the final encoder block using channel-weighted gradients and activations. This map represents where the model is "looking."
Dual-Path Decoding:
1. Main Path: Generates standard segmentation logits.
2. Auxiliary Path: Includes a lightweight refinement pathway and a confidence head that generates a spatial prior to modulate final logits, suppressing distractors.

B. Key Technical Components

Focus Alignment Loss ( $L_{align}$ ):
- Goal: Force the predicted probability mass to align with the XAI attribution field.
- Mechanism: Computes a symmetrized KL divergence and an overlap ratio between the normalized segmentation probabilities and the normalized XAI field. This ensures the model only predicts high probabilities in regions it deems important internally.
Pairwise Consistency Classifier:
- Goal: Enforce anatomical plausibility across consecutive slices.
- Mechanism: A lightweight CNN is trained to distinguish between "realistic" consecutive mask pairs (ground truth) and "unrealistic" non-consecutive pairs. During segmentation training, this classifier acts as a learned regularizer, penalizing predictions that create anatomically impossible transitions between slices.
Anatomy-Aware Losses ( $L_{anat}$ ):
- Goal: Preserve thin structures and topology.
- Mechanism: Combines edge alignment (using differential operators on gradients) and a differentiable skeletonization term (medial axis) to ensure the model respects the thickness and centerline topology of the aneurysm.
Confidence Prior:
- An auxiliary branch predicts a confidence map used to modulate logits at inference, further suppressing low-confidence distractor regions.

C. Training Strategy

The model employs a curriculum learning approach:

Early Phase: Prioritizes slice-to-slice consistency ( $L_{pair}$ ) and confidence coherence to stabilize volumetric coherence.
Late Phase: Shifts emphasis to attribution-probability alignment ( $L_{div}$ ) and anatomical faithfulness ( $L_{anat}$ ) to refine boundary precision.

3. Key Contributions

XAI-Guided Encoder Shaping: Proposes a framework where attribution maps are integrated directly into the training objective to shape encoder representations, rather than just analyzing them post-training.
Focus Alignment Mechanism: Introduces a novel loss function that enforces agreement between the model's internal attention (XAI field) and its output probability mass.
Learned Consistency Regularizer: Instead of hand-crafted penalties, the authors train a classifier on ground-truth mask pairs to learn and enforce realistic slice-to-slice anatomical transitions.
Comprehensive Failure Analysis: Provides a quantitative dissection of model failures (using metrics like Focus Misalignment Index and Boundary Coverage) and demonstrates how XAI guidance corrects them.

4. Experimental Results

The method was evaluated on a clinically curated dataset of 147 CTA scans (9,037 slices), split into a General Test Set and a Complex Test Set (specifically designed to include failure-prone cases).

Performance Metrics (Dice Score / IoU / HD95):

General Test Set: XAI-SAM significantly outperformed baselines (U-Net, U-Net++, MedSAM, SAM).
- Outer Wall: Dice improved from ~89% (SAM baseline) to 97.93%.
- Lumen: Dice improved from ~91.8% to 97.67%.
- Boundary Error (HD95): Reduced drastically from ~4.0mm to 0.19mm for the outer wall.
Complex Test Set: The gains were even more pronounced in difficult cases where baselines failed.
- Outer Wall: XAI-SAM achieved 97.30% Dice and 0.31mm HD95, compared to the SAM baseline's 86.83% Dice and 5.49mm HD95.
- U-Net and U-Net++ showed significant degradation (Dice dropping to ~74-76%) in these complex cases.

Qualitative Findings:

Baseline models (U-Net, SAM) often leaked into adjacent vessels or produced diffuse masks.
XAI-SAM successfully concentrated attention on thin walls and maintained smooth, anatomically consistent transitions across slices.
Ablation studies confirmed that the combination of XAI guidance, anatomy-aware loss, and segmentation refinement is necessary for peak performance.

5. Significance and Conclusion

This paper demonstrates that explainability can be elevated from a diagnostic tool to a guiding principle for training. By explicitly optimizing the encoder's internal focus to align with its predictions, the authors created a segmentation system that is:

More Robust: Highly effective in complex, low-contrast, and failure-prone scenarios where standard models fail.
Anatomically Consistent: Reduces implausible jumps between slices via learned consistency constraints.
Clinically Relevant: Achieves near-perfect boundary alignment (sub-millimeter HD95), which is critical for surgical planning and risk assessment of AAA.

The work suggests a paradigm shift in medical image segmentation: moving beyond simple region-based losses to attribution-driven learning that ensures the model "looks" at the right features before it "predicts" the mask.