Ranking XAI Methods for Head and Neck Cancer Outcome Prediction

This paper presents the first comprehensive evaluation and ranking of 13 explainable AI (XAI) methods across 24 metrics for head and neck cancer outcome prediction, demonstrating that Integrated Gradients and DeepLIFT consistently achieve high performance in faithfulness, complexity, and plausibility on the HECKTOR dataset.

The Gist

Imagine you have a super-smart robot doctor that can look at medical scans of a patient's head and neck and predict how well they will do after treatment. This robot is incredibly accurate, but it has a major flaw: it's a "black box." It gives you the answer, but it won't tell you why it made that decision.

In the real world, doctors can't just trust a robot's gut feeling; they need to know where the robot is looking and what it's thinking. This is where Explainable AI (XAI) comes in. Think of XAI as a translator that tries to explain the robot's thought process to a human doctor.

But here's the problem: There are dozens of different "translators" (XAI methods), and they all tell slightly different stories. Some might point to the tumor, while others might point to the patient's jawbone or a random spot in the air. Until now, no one had really tested which translator was the most honest and reliable for Head and Neck Cancer.

The Big Experiment

The authors of this paper decided to play the role of honest referees. They took 13 different "translators" and put them through a massive test using data from the HECKTOR challenge (a global competition for cancer prediction).

They didn't just ask, "Did it look right?" They tested the translators on four specific qualities, using 24 different rules:

1. Faithfulness (The Truth-Teller): Does the translator actually explain what the robot is thinking, or is it just making things up?
   • Analogy: If the robot says "I'm worried about the tumor," does the translator point to the tumor, or does it point to the patient's ear?
2. Robustness (The Rock): If you nudge the image slightly (like adding a tiny bit of static noise), does the translator keep pointing to the same spot, or does it go crazy and point somewhere else?
   • Analogy: If you whisper a secret to a friend, do they repeat it back correctly, or do they change the story?
3. Complexity (The Concise Speaker): Is the explanation simple and focused, or is it a messy, confusing scribble covering the whole image?
   • Analogy: A good explanation is like a laser pointer; a bad one is like a flashlight beam that lights up the whole room.
4. Plausibility (The Doctor's Eye): Does the explanation look like something a real human doctor would agree with? Does it highlight the actual tumor?
   • Analogy: If a map says "The treasure is here," does the X mark actually land on the gold, or on a tree?

The Results: Who Won?

After running the numbers, the paper found that not all translators are created equal.

• The Champions: Two methods, Integrated Gradients (IG) and DeepLIFT (DL), came out on top. They were the most "faithful" (told the truth about the robot's logic) and the most "plausible" (pointed exactly at the tumors, just like a doctor would).
  • The Metaphor: These two were like the most reliable tour guides. They didn't just show you the scenery; they showed you exactly why the scenery was important, without getting distracted by the clouds or the trees.
• The Runners-Up: Some methods were great at being "stable" (Robustness) but bad at being "truthful." Others were good at being "simple" but missed the tumor entirely.
• The Losers: Some methods (like LIME and OC) were very inconsistent. Sometimes they pointed to the tumor, sometimes to the bone, and sometimes to nothing at all. They were like tour guides who kept changing their minds about where the treasure was buried.

Why This Matters

The paper concludes that you can't just pick a random explanation tool and hope for the best. If you use the wrong one, you might think the AI is looking at the cancer when it's actually looking at a shadow.

The Takeaway:
For Head and Neck Cancer, Integrated Gradients and DeepLIFT are currently the best tools to help doctors trust AI. They act as the most honest interpreters, ensuring that when the AI makes a life-or-death prediction, the "why" behind it is clear, accurate, and medically sensible.

This study is a huge step forward because it moves AI from "magic black box" to "transparent partner," helping doctors feel confident enough to use these powerful tools in real hospitals.

Technical Summary

1. Problem Statement

Head and neck cancer (HNC) patients exhibit significant variability in treatment outcomes despite receiving similar therapies. While deep learning models (e.g., CNNs, Transformers) have improved outcome prediction using PET/CT data, their "black box" nature hinders clinical adoption.

• The Gap: Previous HNC studies have relied on empirical selection of Explainable AI (XAI) techniques (e.g., Grad-CAM) without rigorous quantitative evaluation.
• The Challenge: There is a lack of systematic comparison regarding how well these XAI methods reflect true model reasoning (faithfulness), remain stable under noise (robustness), produce concise maps (complexity), and align with clinical anatomy (plausibility).

2. Methodology

Dataset and Model

• Dataset: The study utilized the HECKTOR 2025 multi-center dataset, comprising 651 patients with CT, PET, and Gross Tumor Volume (GTV) masks. The data was split into 75% training (488 patients) and 25% testing (163 patients).
• Preprocessing: Images were center-cropped to the GTV (192×192×192 mm³), resampled to 2×2×2 mm³, and intensities were normalized.
• Prediction Model: A 3D DenseNet121 architecture was employed to predict Recurrence-Free Survival (RFS). The model was trained using Cox negative log partial likelihood loss, achieving a C-index of 0.66 on the test set.

XAI Evaluation Framework

The authors evaluated 13 post-hoc saliency-based XAI methods categorized into:

1. Perturbation-based: (e.g., OC, LIME, KS) – Masking/altered regions to measure output change.
2. Gradient-based: (e.g., VG, IG, DL) – Backpropagated gradients.
3. CAM-based: (e.g., GC, SC, C+) – Activation maps from convolutional layers.

Evaluation Metrics (24 Total):
The methods were ranked across four dimensions:

• Faithfulness (10 metrics): Agreement with true model reasoning.
• Robustness (5 metrics): Stability under small perturbations/noise.
• Complexity (3 metrics): Conciseness and sparsity of highlighted regions.
• Plausibility (6 metrics): Alignment with clinically relevant anatomy (GTV). Note: This aspect was a novel addition not present in previous benchmarks like LATEC.

Ranking Analysis

Instead of aggregating raw metric values (which have different scales), the authors adopted a ranking-based approach inspired by the LATEC benchmark:

1. Rank all 13 methods for each metric based on mean performance.
2. Calculate the mean, median, and standard deviation of these rankings for each of the four evaluation aspects.
3. Identify methods that consistently achieve top rankings across clinically critical aspects (Faithfulness and Plausibility).

3. Key Results

• Top Performers: Integrated Gradients (IG) and DeepLIFT (DL) emerged as the most robust overall performers.
  • They consistently ranked in the top 3 for Faithfulness, Complexity, and Plausibility.
  • Visual analysis confirmed that IG and DL saliency maps showed the best spatial alignment with the GTV mask, avoiding non-tumor structures (e.g., bone).
• Specific Method Performance:
  • Robustness: Methods like EG, VG, and GC performed best, while IG and DL showed lower robustness (sensitive to noise due to gradient propagation).
  • Plausibility vs. Faithfulness: The study highlighted that high plausibility does not guarantee high faithfulness. For instance, VG ranked high in plausibility (4.7) but low in faithfulness (8.4).
  • Underperformers: Perturbation-based methods (OC, LIME, KS) failed to consistently localize tumors, often producing incomplete or incorrect maps. CAM-based methods (GC, SC, C+) tended to produce global, diffused maps including non-tumor regions.
• Variability: Large standard deviations in rankings across methods confirmed that no single XAI method is universally best; performance is highly dependent on the specific metric and task.

4. Key Contributions

1. First Comprehensive HNC Evaluation: This is the first study to systematically evaluate and rank XAI methods specifically for HNC outcome prediction on the HECKTOR dataset.
2. Expanded Metric Scope: Introduced Plausibility as a critical evaluation dimension (6 metrics), addressing a gap in previous benchmarks (like LATEC) that focused only on faithfulness, robustness, and complexity.
3. Ranking-Based Framework: Demonstrated the utility of ranking-based evaluation to handle the heterogeneity of XAI metric scales, providing a fairer comparison mechanism.
4. Empirical Evidence: Provided concrete evidence that IG and DL are the preferred choices for HNC tasks when balancing clinical plausibility with model reasoning fidelity.

5. Significance and Conclusion

• Clinical Impact: The study provides a roadmap for clinicians and researchers to select trustworthy XAI tools. By prioritizing methods that align with human intuition (plausibility) and true model logic (faithfulness), it bridges the gap between AI performance and clinical interpretability.
• Task-Specificity: The results reinforce that XAI selection must be task-specific. While EG was best in the general LATEC benchmark, IG and DL outperformed it in the specific context of 3D HNC PET/CT prediction.
• Future Directions: The authors suggest future work should focus on adaptive hyperparameter optimization, analyzing metric correlations, and conducting human-in-the-loop evaluations with radiologists to validate quantitative metrics against clinical judgment.

In summary, this paper establishes a rigorous benchmark for XAI in medical imaging, concluding that Integrated Gradients and DeepLIFT currently offer the most reliable explanations for head and neck cancer outcome prediction, provided their sensitivity to noise is managed.