A 2015 CNN Beats GPT-6 Astra and Claude at Melanoma Diagnosis, Study Finds

⏱️ Reading time: 9 min

A dermatologist tested AI melanoma diagnosis on GPT-6 Astra, Claude Fable 5.1, and a neural network published in 2015, using the same photo. The two newer language models said it was a benign wart and that the patient could go home. The old network, much smaller, said melanoma: the correct answer.

📑 En este artículo

1. TL;DR
2. What Is Inception v3
3. What Happened in the AI Melanoma Diagnosis Test
4. Context and History of Neural Networks for Dermatological Diagnosis
5. Technical Details: How a Vision Model Competes at Melanoma Diagnosis
6. Impact and Analysis of AI-Based Skin Lesion Classification
7. What’s Next for AI Skin Cancer Detection
8. Frequently Asked Questions
   1. What is Inception v3 and why was it used for this experiment?
   2. Can GPT-6 Astra and Claude Fable 5.1 diagnose melanoma?
   3. What is MILK10k?
   4. Why do neural networks for dermatological diagnosis beat a large LLM?
   5. Does AI skin cancer detection rule out generic LLMs?
   6. Can you try the Lynch Lab model for free?
9. References

The experiment didn’t stop at one photo. The dermatologist tested all three systems against 200 biopsy-confirmed skin lesions and published the full results on his blog, Lynch Lab.

TL;DR

• A fine-tuned 2015 CNN correctly diagnosed 62% of 200 lesions; the LLMs only got 42%.
• GPT-6 Astra and Claude Fable 5.1 mistook an invasive melanoma for a benign lesion.
• The study used MILK10k, 200 biopsy-confirmed lesions across six categories.
• Claude Fable 5.1 got only 12% of seborrheic keratoses right; the CNN got 52%.
• The test took place in October 2026, and the CNN runs free in the browser via WebGPU.

What Is Inception v3

Inception v3 is a convolutional neural network that Google published in 2015 to classify images through layers of convolution. Fine-tuned on dermatological photographs, it’s the core tool in the AI melanoma diagnosis that a dermatologist compared against two cutting-edge multimodal models on 200 real cases.

The version the dermatologist used isn’t Google’s original network. He started from weights pretrained on generic photos and retrained them completely on 68,011 dermoscopic and clinical images from the ISIC Archive, using a single H100 GPU. None of the 200 test lesions, nor any other photos of the same patient or the same lesion, were part of that training or validation set.

For each lesion, the model receives two images: a clinical photo taken with a regular camera and a dermoscopic image, captured with the illuminated magnifier dermatologists use. It combines both scores, weighting the dermoscopic one more heavily, and chooses among six possible diagnoses.

What Happened in the AI Melanoma Diagnosis Test

The dermatologist behind Lynch Lab set up a level playing field for all three systems. He pulled 200 lesions from the public MILK10k dataset, released by the International Skin Imaging Collaboration in 2025, all with histopathology-confirmed diagnoses. Half were lesions that can’t be cleared without a biopsy: 30 melanomas, 25 basal cell carcinomas, 25 squamous cell carcinomas (including Bowen’s disease), and 20 actinic keratoses. The other half were benign lesions that can be cleared: 50 nevi and 50 seborrheic keratoses.

GPT-6 Astra and Claude Fable 5.1 were given the same two images of each lesion, with no further clinical context. They were asked for the most likely diagnosis among six options, or simply “not sure,” a confidence level, a second choice, and whether the lesion could be cleared without a biopsy or follow-up. Neither model declined to answer on any case.

The aggregate result was clear. Inception v3 got 125 of 200 diagnoses right (62%). GPT-6 Astra got 84 right (42%) and Claude Fable 5.1 got 85 right (42%). A network fine-tuned in an afternoon of GPU time beat the two most capable multimodal models of October 2026 by 20 percentage points.

The most concerning case is an invasive melanoma on the ear, less than 0.8 mm thick. Both images showed little pigmentation. GPT-6 Astra and Claude Fable 5.1 called it a seborrheic keratosis and recommended clearing the patient without a biopsy. Inception v3, looking at the dermoscopic image, correctly identified it as melanoma.

In another case, an equally thin melanoma with little pigment, GPT-6 Astra classified it as basal cell carcinoma with 95% confidence. Claude Fable 5.1 and Inception v3 did say melanoma. There was also a case where all three got it wrong: a benign seborrheic keratosis that both LLMs mistook for actinic keratosis and the CNN mistook for squamous cell carcinoma. There, the error came cheaper, since none of the three would have cleared the lesion.

💭 Key point: All 200 lesions, including the benign ones, had been biopsied because a doctor had suspected something earlier. The experiment measures the worst-case scenario, not the typical mole that never makes it to a clinic.

Context and History of Neural Networks for Dermatological Diagnosis

Inception v3 isn’t new. Google published it in 2015 as an architectural improvement over its own original Inception network, designed to compete in ImageNet, a competition for classifying generic photos (cats, cars, instruments), not skin. A decade of use made it one of the most proven architectures for fine-tuning on small vision tasks, because there’s nothing new to invent: you just retrain its final layers on your own dataset and you’re done.

Multimodal language models came later with a different promise: a single system that understands text, images, and clinical context at once, without building a 68,000-photo dataset or paying for a dedicated GPU. As of October 2026, GPT-6 Astra and Claude Fable 5.1 are two of the most capable multimodal models available to the public. But neither was trained specifically to tell apart six types of skin lesion: they learned dermatology from the same internet mix they learned everything else from.

Technical Details: How a Vision Model Competes at Melanoma Diagnosis

Inception v3’s advantage doesn’t come from being bigger. It comes from having a tiny response space, just six classes, where every training example reinforces exactly that decision boundary. A generalist multimodal model spreads its capacity across thousands of possible tasks. When it sees a dermoscopic image, it activates representations learned from photos across the internet in general, not from a dataset curated by dermatologists.

flowchart TD
A["Clinical photo"] --> C["Classification model"]
B["Dermoscopic image"] --> C
C --> D[("Diagnosis among 6 categories")]
subgraph Decision
C
D
end

Fine-tuning also matters for the type of error it produces. By retraining every layer of Inception v3 on clinical and dermoscopic images, the gradient directly adjusts the filters that detect irregular borders, asymmetry, and color variation: the signals a dermatologist looks for by eye in dermoscopy. A multimodal LLM, by contrast, describes the image with the same attention process it uses to read a screenshot or a meme. There’s no dedicated layer to weigh the asymmetry of a mole.

This doesn’t mean LLMs are useless in healthcare. GPT-6 Astra got 100% of basal cell carcinomas right, better than the CNN. The difference lies in the cost of errors: missing a basal cell carcinoma almost never kills anyone if it’s caught later, but missing a melanoma can.

Impact and Analysis of AI-Based Skin Lesion Classification

For AI melanoma diagnosis to be clinically useful, it needs specialization, not just parameter count. The same pattern has already shown up in radiology, where models trained to read chest X-rays outperform generalists in sensitivity to rare findings, even if they lose out in natural conversation.

Retraining a network like Inception v3 for a new task isn’t expensive. It ran on a single H100 GPU and used a public dataset: nothing a small data team couldn’t replicate. Building a multimodal LLM at the scale of GPT-6 Astra or Claude Fable 5.1, on the other hand, requires the infrastructure of a frontier lab. The paradox is that the more technically accessible tool turned out to be the most accurate at the highest-risk task.

What’s Next for AI Skin Cancer Detection

The dermatologist released a public demo of the model that runs entirely in the browser via WebGPU, without uploading any photo to a server, and lets you test random lesions from the same set of 200 cases. It’s the most direct way to see the difference between a specialized classifier and a generic LLM without installing anything.

What still doesn’t exist is a comparable version trained specifically for this task inside GPT-6 Astra or Claude Fable 5.1. Neither lab had announced plans for specialized medical fine-tuning in dermatology as of this writing. Until that happens, the practical recommendation for a clinic doesn’t change: a small fine-tuned model remains the safest option for medical image classification, and a multimodal LLM works as conversational support, not as a biopsy reader.

Try it yourself: open the CNN demo on the Lynch Lab blog and compare your own judgment against the three classifications before trusting a chatbot to read a skin lesion.

Frequently Asked Questions

What is Inception v3 and why was it used for this experiment?

It’s a convolutional neural network that Google published in 2015. The Lynch Lab dermatologist retrained all its layers on 68,011 dermatological images from the ISIC Archive, using a single H100 GPU, to turn it into a classifier for six types of skin lesion.

Can GPT-6 Astra and Claude Fable 5.1 diagnose melanoma?

They can offer an opinion, but in this test they got the correct diagnosis in only 42% of 200 biopsied lesions, versus 62% for the specialized CNN. Both specifically failed to distinguish melanoma from benign seborrheic keratosis.

What is MILK10k?

It’s a public dataset of skin lesions released in 2025 by the International Skin Imaging Collaboration, with clinical and dermoscopic photos of each case, confirmed by biopsy.

Why do neural networks for dermatological diagnosis beat a large LLM?

Because their response space has only six classes, and all their training reinforces that decision boundary. A fine-tuned melanoma vision model adjusts specific filters for asymmetry and irregular borders; a generalist LLM spreads its capacity across thousands of different tasks.

Does AI skin cancer detection rule out generic LLMs?

Not entirely. GPT-6 Astra got 100% of the study’s basal cell carcinomas right, better than the CNN. The point is that the high-risk task needs a specialized model, not a generic one, even if the generic one is bigger.

Can you try the Lynch Lab model for free?

Yes. It runs entirely in the browser via WebGPU, without uploading photos to any server, and lets you pick random lesions from the same set of 200 cases used in the test.

References

• Lynch Lab: the original article with the full methodology and the browser-based CNN demo.
• ISIC Archive: the repository hosting the MILK10k dataset used in the test.
• arXiv: the original Inception v3 paper published by Google in 2015.
• Wikipedia: general clinical context on melanoma and its diagnosis.

📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day.

Featured image: Foto de CDC en Unsplash