Source-linked AI summary
Dermatologist-like explainable AI enhances trust and confidence in diagnosing melanoma
Tirtha Chanda, Katja Hauser, Sarah Hobelsberger, Tabea-Clara Bucher, Carina Nogueira Garcia, Christoph Wies, Harald Kittler, Philipp Tschandl, Cristian Navarrete-Dechent, Sebastian Podlipnik, Emmanouil Chousakos, Iva Crnaric, Jovana Majstorovic, Linda Alhajwan, Tanya Foreman, Sandra Peternel, Sergei Sarap, İrem Özdemir, Raymond L. Barnhill, Mar Llamas Velasco, Gabriela Poch, Sören Korsing, Wiebke Sondermann, Frank Friedrich Gellrich, Markus V. Heppt, Michael Erdmann, Sebastian Haferkamp, Konstantin Drexler, Matthias Goebeler, Bastian Schilling, Jochen S. Utikal, Kamran Ghoreschi, Stefan Fröhling, Eva Krieghoff-Henning, Titus J. Brinker
TL;DR
The paper addresses limited transparency and difficult-to-interpret explanations in AI-assisted melanoma diagnosis. It develops a multimodal XAI with human-relevant text and region explanations and evaluates it in a three-phase reader study. The system preserved diagnostic performance, significantly increased clinicians’ confidence and trust, and aligned its explanations with clinicians’ reasoning, while its accuracy advantage over conventional AI was not significant.
Problem
AI systems can improve melanoma diagnosis, but opaque decisions and imprecisely located explanations limit clinicians’ ability to interpret and accept them.
Method
The study developed a multimodal XAI using dermatologist-annotated melanoma and nevus features and regions, then evaluated it in a three-phase reader study.
Results
Clinicians’ confidence and trust significantly increased with XAI, while diagnostic accuracy did not significantly improve beyond AI support alone and the XAI matched baseline diagnostic performance.
Takeaways & Limitations
Human-relevant explanations can support clinicians’ confidence and trust while retaining diagnostic performance in melanoma AI assistance.
Takeaways & Limitations
The system was tested under artificial conditions, intentionally favored human-like explanations over potentially non-human patterns, and did not address domain shift.
Abstract
from arXiv · showhide
Although artificial intelligence (AI) systems have been shown to improve the accuracy of initial melanoma diagnosis, the lack of transparency in how these systems identify melanoma poses severe obstacles to user acceptance. Explainable artificial intelligence (XAI) methods can help to increase transparency, but most XAI methods are unable to produce precisely located domain-specific explanations, making the explanations difficult to interpret. Moreover, the impact of XAI methods on dermatologists has not yet been evaluated. Extending on two existing classifiers, we developed an XAI system that produces text and region based explanations that are easily interpretable by dermatologists alongside its differential diagnoses of melanomas and nevi. To evaluate this system, we conducted a three-part reader study to assess its impact on clinicians' diagnostic accuracy, confidence, and trust in the XAI-support. We showed that our XAI's explanations were highly aligned with clinicians' explanations and that both the clinicians' trust in the support system and their confidence in their diagnoses were significantly increased when using our XAI compared to using a conventional AI system. The clinicians' diagnostic accuracy was numerically, albeit not significantly, increased. This work demonstrates that clinicians are willing to adopt such an XAI system, motivating their future use in the clinic.
Results
The XAI matched baseline diagnostic accuracy while providing explanations aligned with clinicians’ ontological features and lesion regions. It significantly increased clinicians’ confidence and trust, but did not significantly improve diagnostic accuracy beyond AI support alone.
- Diagnostic performance: 81% balanced accuracy for the XAI versus 80% for the baseline classifier, with overlapping confidence intervals.The XAI maintained diagnostic performance while adding interpretability.
- Diagnostic performance: AI support increased mean balanced accuracy from 66.2% without AI to 72.3% with AI support (P<0.0001).This comparison isolates the benefit of AI support over no AI support.
- Diagnostic performance: 73.2% mean balanced accuracy with XAI was not significantly higher than accuracy with AI support alone (P=0.34).Accuracy improved for 52 participants, decreased for 34, and was unchanged for 30.
- Confidence: 12.25% mean confidence increase with XAI relative to AI support alone was statistically significant (P<0.0001).Clinicians’ confidence was also slightly higher for high-confidence than uncertain AI predictions.
- Trust: 17.52% mean trust increase with XAI relative to AI support alone was statistically significant (P<0.0001).Trust was slightly correlated with explanation overlap when clinicians and AI agreed, especially for melanoma diagnoses.
Classifier design
The classifier predicts dermatologist-defined lesion characteristics rather than diagnoses directly, then infers melanoma or nevus while using attention guidance to localize relevant features. Training combines characteristic classification with ROI-based attention supervision and calibrated confidence outputs.
- The classifier predicts visual characteristics from dermoscopic images and infers melanoma or nevus from those predictions.
- A classification component predicts characteristics, while a guided attention component helps localize the corresponding features.
- The attention component uses dermatologist-annotated ROIs to guide the model toward features used by dermatologists and produce more meaningful explanations.
- Training combines cross-entropy for characteristic prediction with a Dice-based attention loss weighted by λC=1 and λA=10.
- The model uses a pretrained ResNet50 with balanced sampling and image augmentation to improve generalizability.
- Temperature scaling calibrates output probabilities so displayed confidence more accurately reflects model confidence.
Classifier performance testing
Classifier testing evaluated balanced accuracy and lesion-focused attention, while explanation overlap was quantified with Sørensen-Dice similarity coefficients. These measures assess predictive performance, spatial focus, and agreement between model and clinician explanations.
- Balanced accuracy evaluated classifier performance on a held-out test set, while sensitivity and specificity were determined on a validation set.
- The attribution ratio compares mean Grad-CAM activation inside the lesion with activation in the surrounding skin.
- Sørensen-Dice similarity coefficients quantified overlap between XAI-predicted and clinician-selected explanations using true positives, false positives, and false negatives.
- The same Dice-based comparison was used for ROI overlap, with soft image masks and a smoothing term ε.
Design of the explanations
The explanation design was developed with dermatologists and combines text, localized visual regions, and confidence information. Polygon ROIs and confidence labels were chosen to make feature-level predictions easier to interpret clinically.
- Two board-certified dermatologists helped design an explanation scheme combining visual and text-based components with classification confidence.
- Because conventional heatmaps can obscure lesions, the system displays polygon-shaped ROIs over the top 20th percentile attribution values.
- The system communicates uncertainty by labeling detected characteristics as either “strong evidence” or “some evidence.”
- Localized explanations show ROIs for characteristics whose logits exceed 0.7, or for the most certain characteristic when none exceeds that threshold.
- The three-part reader study took place between July and December 2022.
Participants
The reader study involved international dermatology clinicians diagnosing lesions across three phases with no AI, conventional AI, and XAI support. Participants reported diagnoses, confidence, and—when AI support was provided—trust, while analyses used paired and nonparametric tests.
- Participants: The study recruited 120 international dermatology clinicians for phase 1, with 116 completing the complete study.
- Phase 1: Phase 1 asked clinicians to diagnose 15 lesions, select explanatory characteristics, annotate image regions, and rate diagnostic confidence.
- Phase 1: The lesion sets comprised balanced melanoma and nevus images, with one repeated image per group and 196 unique test images overall.
- Phase 1: Clinicians could select “nevus (leave in),” “nevus (excise),” or “melanoma,” with both nevus options treated as nevus for accuracy evaluation.
- Phases 2–3: Phase 2 provided conventional AI diagnoses, while phase 3 provided XAI support for the same lesions without informing participants of the repeated cases.
- Phases 2–3: Confidence and trust were rated on 1–10 Likert scales in the AI-supported phases, and 116 participants completed phase 3.
- Analysis: The analysis used paired t tests, Mann–Whitney U tests, Wilcoxon signed-rank tests, and bootstrap confidence intervals.
- Robustness: Ruler artefact sensitivity was compared using mean absolute output changes for the XAI and baseline classifier.