A new artificial intelligence tool called Dermoscopy Explainable Intelligence (DEXI) correctly identified almost 94% of skin cancers and ruled out about 75% of non-cancerous lesions in a Mayo Clinic study published in Mayo Clinic Proceedings. The system assigns lesions a score from zero to 10 to help dermatologists determine whether to monitor a spot or perform a biopsy.
In Plain English: The Clinical Takeaway
- What DEXI does: It analyzes photographs of skin spots and gives doctors a risk score from 0 to 10 along with details on asymmetry, border, and color.
- Accuracy rates: The tool correctly flagged nearly 94% of actual skin cancers and accurately dismissed about 75% of harmless lesions in testing.
- Clinical limits: Mayo Clinic researchers stress the tool must not be used alone, as it does not account for patient age, medical history, or current medications.
Mayo Clinic Evaluates New Scoring System for Skin Lesions
Researchers tested DEXI on 261 lesions found on 178 patients who underwent total-body photography. The system was used to assist clinicians in evaluating spots. For noncancerous lesions, the average DEXI score was 2.5. For skin cancers, the average jumped to 7.8, while melanomas averaged 6.0.
As Microsoft Build concludes, Mayo Clinic's collaboration with @Microsoft marks an important step in the ongoing development of AI tools designed specifically for healthcare. During the conference, Dr. @GFarrugiaMD, president and CEO of Mayo Clinic, discussed the collaboration,… pic.twitter.com/GRDA6ynipT
Based on these outcomes, the research team identified 3.0 as a potential threshold for potential malignancy. A score at or above this level gives dermatologists quantitative data to weigh when deciding between regular monitoring or an immediate tissue sample extraction. Dr. Jenny Link, a dermatologist at the Mayo Clinic in Rochester, Minnesota, noted that prior to the study, a score lacked clear clinical direction, but now serves as a piece of data to inform clinical judgment.
Older Specialized Models Outperform Frontier AI in Comparative Tests
While imaging tools like DEXI assist specialized physicians, a separate test revealed that older, purpose-built networks can outperform modern general-purpose AI. Dr. Magnus Lynch, principal investigator at the Lynch Institute at King’s College London, compared frontier language and vision models against Google’s Inception v3—a convolutional neural network released in 2016.
When tested on 200 skin lesion images split evenly between treatable and harmless cases, Inception v3—fine-tuned for 20 minutes on 68,011 skin images—correctly identified diagnoses in 125 cases (62.5%). By comparison, GPT-6 Astra and Claude Fable 5.1 correctly identified 84 and 85 cases respectively. Lynch pointed out that specialized networks using older, smaller architectures show superior lesion recognition when properly trained on specific medical data, whereas unadjusted general-purpose models struggle with nuanced dermatological tasks.
| AI Model | Architecture / Release | Correct Diagnoses (out of 200) | Melanoma Correctly Identified |
|---|---|---|---|
| Inception v3 | Specialized CNN (2016, fine-tuned) | 125 (62.5%) | 13 / 30 |
| Claude Fable 5.1 | General-Purpose LLM (2026) | 85 (42.5%) | 11 / 30 |
| GPT-6 Astra | General-Purpose LLM (2026) | 84 (42.0%) | 8 / 30 |
Human Experts Maintain the Edge in Comprehensive Diagnostic Trials
Algorithm performance also shifts when evaluated against human specialists handling complex real-world data. A study published in JAMA Dermatology pitted three diagnostic AI systems against 652 physicians using 1,117 standardized cases featuring uncommon lesions and uneven image qualities.
Experienced dermatologists achieved a 74.2% accuracy rate, outperforming the AI systems, which ranged from 56.7% to 72.2% accuracy. While advanced models matched physicians with three to 10 years of experience, they fell short against specialists with more than a decade in practice. Researchers noted that AI often falters when processing blurry photos, rare cancers, or unusual moles, and its strength in flagging cancer does not always translate to naming the exact diagnosis among multiple possibilities.
Contraindications & When to Consult a Doctor
Patients must understand that AI imaging tools and smartphone apps are designed strictly as auxiliary screening aids rather than definitive diagnostic replacements. Individuals should avoid relying solely on automated algorithms for self-diagnosis, especially since tools like DEXI do not factor in patient age, personal history of melanoma, or current pharmacological treatments.
Professional medical intervention is mandatory if you notice a new, rapidly changing, bleeding, or otherwise unusual skin lesion. Anyone spending extensive time in high-UV environments should schedule an annual full-body skin check with a board-certified dermatologist.
References
- Mayo Clinic Proceedings — AI tool may support dermatologists in assessing skin lesions for cancer risk
- JAMA Dermatology — Comparative Performance of AI Systems and Physicians in Skin Cancer Diagnosis
- Gigazine — A 10-year-old CNN outperforms frontier LLMs for skin cancer diagnosis
- Digital Today — Older Google model beats latest GPT and Claude in skin cancer test
- Miami Herald — Experienced dermatologists came out ahead of AI, study found
Keep reading
- Blood tests can identify bowel disease signs before onset
- New Chest Pain and Gas During Esophagitis Recovery Explained
- Cardinals QB Carson Beck Injury Update: Mike LaFleur Evaluates Rookie’s Status (news-usa.today)
- Alan Dysert, ‘All My Children’ Actor, Dies at 75 After Cancer Battle (world-today-journal.com)
