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,Alexander Salava,Alexander Thiem,Alexandris Dimitrios,Amr Mohammad Ammar,Ana Sanader Vučemilović,Andrea Miyuki Yoshimura,Andzelka Ilieva,Anja Gesierich,Antonia Reimer-Taschenbrecker,Antonios G. A. Kolios,Arturs Kalva,Arzu Ferhatosmanoğlu,Aude Beyens,Claudia Pföhler,Dilara Ilhan Erdil,Dobrila Jovanovic,Emoke Racz,Falk G. Bechara,Federico Vaccaro,Florentia Dimitriou,Gunel Rasulova,Hulya Cenk,Irem Yanatma,Isabel Kolm,Isabelle Hoorens,Iskra Petrovska Sheshova,Ivana Jocic,Jana Knuever,Janik Fleißner,Janis Raphael Thamm,Johan Dahlberg,Juan José Lluch-Galcerá,Juan Sebastián Andreani Figueroa,Julia Holzgruber,Julia Welzel,Katerina Damevska,Kristine Elisabeth Mayer,Lara Valeska Maul,Laura Garzona-Navas,Laura Isabell Bley,Laurenz Schmitt,Lena Reipen,Lidia Shafik,Lidija Petrovska,Linda Golle,Luise Jopen,Magda Gogilidze,Maria Rosa Burg,Martha Alejandra Morales-Sánchez,Martyna Sławińska,Miriam Mengoni,Miroslav Dragolov,Nicolás Iglesias-Pena,Nina Booken,Nkechi Anne Enechukwu,Oana-Diana Persa,Olumayowa Abimbola Oninla,Panagiota Theofilogiannakou,Paula Kage,Roque Rafael Oliveira Neto,Rosario Peralta,Rym Afiouni,Sandra Schuh,Saskia Schnabl-Scheu,Seçil Vural,Sharon Hudson,Sonia Rodriguez Saa,Sören Hartmann,Stefana Damevska,Stefanie Finck,Stephan Alexander Braun,Tim Hartmann,Tobias Welponer,Tomica Sotirovski,Vanda Bondare-Ansberga,Verena Ahlgrimm-Siess,Verena Gerlinde Frings,Viktor Simeonovski,Zorica Zafirovik,Julia-Tatjana Maul,Saskia Lehr,Marion Wobser,Dirk Debus,Hassan Riad,Manuel P. Pereira,Zsuzsanna Lengyel,Alise Balcere,Amalia Tsakiri,Ralph P. Braun,
DOI: https://doi.org/10.1038/s41467-023-43095-4
IF: 16.6
2024-01-15
Nature Communications
Abstract:Abstract Artificial intelligence (AI) systems have been shown to help dermatologists diagnose melanoma more accurately, however they lack transparency, hindering user acceptance. Explainable AI (XAI) methods can help to increase transparency, yet often lack precise, domain-specific explanations. Moreover, the impact of XAI methods on dermatologists’ decisions has not yet been evaluated. Building upon previous research, we introduce an XAI system that provides precise and domain-specific explanations alongside its differential diagnoses of melanomas and nevi. Through a three-phase study, we assess its impact on dermatologists’ diagnostic accuracy, diagnostic confidence, and trust in the XAI-support. Our results show strong alignment between XAI and dermatologist explanations. We also show that dermatologists’ confidence in their diagnoses, and their trust in the support system significantly increase with XAI compared to conventional AI. This study highlights dermatologists’ willingness to adopt such XAI systems, promoting future use in the clinic.
multidisciplinary sciences
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is how to improve the accuracy and confidence of dermatologists in diagnosing melanoma through an Explainable AI (XAI) system. Specifically, the research aims to: 1. **Improve transparency**: Although existing deep neural networks (DNNs) can help dermatologists diagnose melanoma and moles more accurately, these models lack transparency, resulting in low user acceptance. This paper introduces a new XAI system that not only provides differential diagnosis but also provides precise and domain - specific explanations to increase transparency. 2. **Evaluate the impact on diagnosis**: Through a three - stage study, evaluate the impact of this XAI system on the diagnostic accuracy, diagnostic confidence, and trust in the support system of dermatologists. 3. **Promote clinical application**: The research results show that the XAI system can significantly improve dermatologists' confidence in their diagnosis and trust in the support system, thus promoting the application of such systems in the clinic. ### Research background - **The importance of melanoma**: Melanoma is the cause of most skin - cancer - related deaths worldwide. Early detection and excision are crucial to ensure the best prognosis for patients. - **Limitations of existing AI systems**: Although AI - based support systems can help dermatologists diagnose melanoma and moles more accurately, these systems lack transparency and it is difficult to explain their decision - making processes, which hinders their wide application in clinical practice. - **Regulatory requirements**: According to the General Data Protection Regulation (GDPR), all algorithm - based decisions must be explainable, and existing DNNs usually cannot meet this requirement. ### Research methods - **Develop an XAI system**: Researchers developed a multimodal XAI system that can locate and explain features in dermoscopic images, making its explanations consistent with the diagnostic perspective of dermatologists. - **Create a dataset**: To train and evaluate this XAI system, researchers created an expert - annotated dataset. - **Three - stage reader study**: Quantify the impact of the XAI system on clinicians through a three - stage reader study involving 116 international participants. ### Main findings - **Diagnostic accuracy**: Although the XAI system did not significantly improve the diagnostic accuracy of dermatologists, it significantly improved their confidence in their own diagnosis and trust in the support system. - **Explanation consistency**: The explanations of the XAI system are highly consistent with those of dermatologists, especially at the lesion - level explanations. - **Potential for clinical application**: The study shows that dermatologists are willing to adopt this XAI system, laying the foundation for future clinical applications. ### Conclusion By introducing a new XAI system, this study successfully improved dermatologists' trust and confidence in AI - support systems, although it did not significantly improve diagnostic accuracy. This provides an important step for further clinical application of XAI systems to improve patient treatment outcomes.