💻 computer science

When AI and Experts Agree on Error: Intrinsic Ambiguity in Dermatoscopic Images

This study reveals that certain dermatoscopic images possess intrinsic visual ambiguity that causes systematic diagnostic failures in both AI models and human experts, as evidenced by a significant collapse in agreement rates and inter-rater reliability when compared to control cases.

Original authors: Loris Cino, Pier Luigi Mazzeo, Alessandro Martella, Giulia Radi, Renato Rossi, Cosimo Distante

Published 2026-04-02

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Loris Cino, Pier Luigi Mazzeo, Alessandro Martella, Giulia Radi, Renato Rossi, Cosimo Distante

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

The Big Idea: The "Impossible" Picture

Imagine you are in a classroom with two groups:

The Super-Computers (AI): A team of five different, highly advanced AI models trained to look at pictures of skin moles and tell you if they are dangerous (cancer) or safe.
The Super-Experts (Doctors): A panel of six veteran dermatologists with 20+ years of experience.

Usually, we think of AI and Doctors as rivals. We ask, "Who is better?" But this study asked a different question: "What happens when both the AI and the Doctors get the same picture wrong?"

The researchers found a specific group of skin pictures where everyone failed. The AI guessed wrong, and the human experts also guessed wrong.

The Analogy: The Foggy Window

Think of these difficult skin images like a window covered in thick fog.

The AI tries to look through the fog and sees shapes that aren't there.
The Doctor tries to look through the same fog and also can't see clearly.

The study discovered that when the "fog" (poor image quality) is too thick, even the smartest computer and the most experienced doctor can't tell the difference between a harmless mole and a dangerous one. They aren't failing because they are "stupid"; they are failing because the picture itself is broken or unclear.

The Three Main Discoveries

1. The "Shared Confusion" Zone

The researchers found that there are certain images where the AI makes mistakes more often than random chance would predict. It's not just one AI getting confused; it's all of them.

The Metaphor: Imagine a group of five different weather forecasters. If they all predict "sunny" when it's actually pouring rain, it's not a coincidence. It means the data they are looking at is misleading.
The Result: When the AI got these specific images wrong, the human doctors also got them wrong at a much higher rate than usual. The doctors' agreement with each other dropped, and their accuracy plummeted.

2. The "Blurry Photo" Culprit

Why did everyone fail? The study found the main culprit: Bad Image Quality.

Many of the "impossible" pictures were blurry, out of focus, or had hair covering the spot.
The Metaphor: It's like asking someone to identify a specific car model from a photo taken through a dirty, foggy windshield at night. No matter how good the driver (Doctor) or the camera (AI) is, the photo is too blurry to tell a Ford from a Toyota.
The Fix: The team built a "Blur Detector" (using math tools like Fourier transforms) that can automatically spot these bad photos and throw them out before they confuse the system.

3. The "Patient Context" Surprise

Doctors usually use more than just a picture to diagnose. They ask: "How old is the patient?" "Where on the body is this?" "Is it a man or a woman?"

The researchers tried to feed this extra information into the AI to help it.
The Surprise: It didn't help. Even with the patient's age and gender, the AI still struggled with the blurry, difficult images.
The Lesson: You can't fix a bad photo by adding more text. If the picture is unclear, knowing the patient's age won't magically make the cancer visible.

Why Does This Matter?

This paper changes the conversation about AI in medicine.

Old View: "AI is failing because it's not smart enough. We need to make better AI."
New View: "AI is failing because the data is messy. Sometimes, the picture is just too hard for anyone to solve."

The Takeaway:
Before we blame the AI for making mistakes, we need to check the "camera." If the image is blurry or low-quality, we shouldn't expect the AI (or the doctor) to be perfect. The solution isn't just smarter algorithms; it's better photography and cleaner data.

Summary in One Sentence

This study shows that when AI and human experts both get a skin diagnosis wrong, it's often because the photo is too blurry or unclear to be solved by anyone, proving that bad data is a bigger problem than bad algorithms.

1. Problem Statement

The integration of Convolutional Neural Networks (CNNs) into dermatological diagnosis has shown promise, yet clinical adoption remains limited. A critical gap in existing literature is the assumption that AI failures are solely due to algorithmic limitations or training biases. This study posits that a significant subset of diagnostic errors arises from intrinsic ambiguity within the images themselves—cases where both AI models and human experts struggle simultaneously.

The research addresses three specific challenges:

Systematic Misclassification: Identifying images that are consistently misclassified by multiple diverse CNN architectures.
Human-AI Correlation: Determining if these "difficult" images also cause diagnostic failures for expert dermatologists.
Data Quality: Investigating whether image artifacts (e.g., blur, low resolution) contribute to these shared errors.

2. Methodology

A. Dataset and Models

Dataset: The study utilized the ISIC 2019 dataset, containing dermatoscopic images across 8 diagnostic classes (e.g., Melanoma, Basal Cell Carcinoma, Nevus).
AI Models: Five distinct pretrained CNN architectures were employed: ResNeXt-50, ResNet-152, EfficientNet-B4, EfficientNet-B5, and EfficientNet-B6.
Training Protocol: Models were pretrained on ImageNet and fine-tuned on ISIC 2019 using 5-fold cross-validation, resulting in 25 total trained models.

B. Identification of "Difficult" Images

The authors isolated a subset of images that were consistently misclassified by all five CNN models.

Statistical Validation: A stratified permutation test (100,000 iterations) was conducted to verify if the number of jointly misclassified images exceeded random chance.
Null Hypothesis ( $H_0$ ): All images have an equal probability of being misclassified; errors are independent of image identity.
Result: The observed number of shared errors (823 images) was statistically impossible under the null hypothesis ( $p \approx 0$ ), confirming the existence of a systematic "difficult" subset.

C. Human Expert Evaluation

Participants: Six expert dermatologists (each with >20 years of experience).
Procedure: Experts evaluated two groups via a custom web platform:
1. Difficult Group: 190 images misclassified by all CNNs.
2. Control Group: 80 images correctly classified by all CNNs.
Blind Protocol: Ground-truth labels were hidden. Experts provided diagnoses based on visual inspection and patient metadata.
Metrics: Agreement was measured using Cohen's kappa (expert vs. ground truth) and Fleiss' kappa (inter-rater reliability among experts).

D. Image Quality Analysis

To determine if low quality drove these errors, the authors applied three automated blur-detection techniques to the dataset:

Laplacian Filter: Detects edges and rapid intensity changes.
Fourier Transform: Analyzes frequency components (blur attenuates high frequencies).
Wavelet Transform: Decomposes signals across scales to detect high-frequency loss.

A unified "blur score" was derived to systematically identify and filter low-quality images.

3. Key Results

A. Statistical Significance of Shared Errors

The permutation test confirmed that the 823 images misclassified by all models were not random outliers but a distinct, inherently difficult subset ( $p < 0.05$ ).

B. Collapse in Human Diagnostic Performance

When experts evaluated the "Difficult" group, their performance degraded significantly compared to the "Control" group:

Agreement with Ground Truth (Cohen's Kappa):
- Control Group: 0.61 (Substantial agreement).
- Difficult Group: 0.08 (Negligible agreement, barely better than chance).
Inter-Rater Reliability (Fleiss' Kappa):
- Control Group: 0.456 (Moderate concordance).
- Difficult Group: 0.275 (Modest agreement).
Specific Class Failures:
- Basal Cell Carcinoma (BCC): Experts failed to identify a single BCC case correctly in the difficult group (Sensitivity = 0.00).
- Melanoma (MEL): Sensitivity dropped from 0.90 (Control) to 0.53 (Difficult).
- VASC (Vascular lesions): Both AI and humans maintained high accuracy, suggesting visual inspection is sufficient for this class.

C. Impact of Image Quality

Blur Correlation: A significant portion of the "difficult" images were identified as low-quality or blurred.
Automated Filtering: The combined Laplacian/Wavelet approach successfully identified blurred images. A threshold of -0.7 removed 61% of blurred images while retaining 74% of high-quality images.
Duplicates: The analysis also revealed duplicate images within the dataset (e.g., identical lesions labeled differently or appearing multiple times), further complicating evaluation.

D. Meta-Classifier Performance (Appendix C)

The authors tested ensemble methods and meta-classifiers (Random Forest, XGBoost, etc.) using CNN predictions and patient data (age, gender, location).

Best Performance: A Random Forest meta-classifier achieved a Balanced Accuracy of 89.75%, outperforming single CNNs (max 85.36%).
Patient Data: Surprisingly, adding patient metadata (age, gender, location) did not significantly improve predictive accuracy, suggesting that for these specific datasets, visual features dominate demographic cues.

4. Key Contributions

Evidence of Intrinsic Ambiguity: The study provides statistical proof that a subset of dermatoscopic images is inherently ambiguous, causing simultaneous failure in both state-of-the-art AI and human experts.
Human-AI Error Correlation: It demonstrates that AI misclassification is a strong predictor of human diagnostic uncertainty, challenging the notion that AI errors are purely algorithmic.
Quality-Centric Analysis: The research highlights that image artifacts (blur, occlusion) are primary drivers of these shared errors, advocating for rigorous data quality control.
Open Science: The authors released the full dataset of "difficult" images, code, and trained models to promote reproducibility and further research into ambiguous cases.

5. Significance and Implications

Clinical Reality Check: The findings suggest that the "ceiling" for diagnostic accuracy in certain cases is limited by the image quality and visual ambiguity, not just the sophistication of the algorithm.
Dataset Curation: The study underscores the urgent need for standardized, automated quality-assessment pipelines in dermatology datasets to filter out blurred or duplicate images before training.
Future AI Development: Rather than simply stacking more models, future research should focus on:
- Developing models specifically for "hard" cases.
- Integrating multi-modal data (e.g., physical examination, palpation) rather than relying solely on static images.
- Improving image acquisition protocols to ensure high-resolution, in-focus inputs.
Reframing Evaluation: The paper argues for shifting the benchmark from "AI vs. Human" to understanding "When and Why" both fail, which is crucial for safe clinical deployment.

In conclusion, this paper shifts the narrative from AI superiority to shared diagnostic limitations, emphasizing that improving dataset quality and understanding intrinsic image ambiguity are prerequisites for the next generation of reliable dermatological AI.