Ten-Year-Old Inception Convolutional Neural Network Outperforms Frontier LLMs on Skin Cancer Diagnosis
Dr. Magnus Lynch benchmarked Inception v3 against GPT-6 Astra and Claude Fable 5.1 across 200 biopsy-verified skin lesions. The 2015 convolutional architecture scored 62% accuracy against 42% for both frontier language models, while running locally in-browser via WebGPU.
Why it matters
Stop routing domain computer vision tasks to expensive cloud multimodal foundation models when lightweight, fine-tuned convolutional networks deliver higher accuracy at zero API cost.
TL;DR
- 01Inception v3 fine-tuned in 20 minutes achieved 62% diagnostic accuracy versus 42% for GPT-6 Astra and Claude Fable 5.1.
- 02Generalist LLMs exhibited severe classification biases, misidentifying up to 46% of melanomas as basal cell carcinoma.
- 03The specialized CNN architecture runs fully client-side on mobile devices via WebGPU without data transmission.
Key facts
- Inception v3 Diagnostic Accuracy
- 62% (125/200 lesions)
- Claude Fable 5.1 Diagnostic Accuracy
- 42.5% (85/200 lesions)
- GPT-6 Astra Diagnostic Accuracy
- 42% (84/200 lesions)
- Fine-Tuning Hardware & Time
- 20 minutes on 1x NVIDIA H100 GPU
- Training Dataset Size
- 68,011 ISIC Archive images
- Second-Choice Accuracy (CNN vs LLMs)
- 173/200 (CNN) vs 138/200 (Claude) and 134/200 (GPT-6)
Benchmark Methodology and Results
Dr. Magnus Lynch compared three architectures across 200 histopathologically verified skin lesions selected from the 2025 MILK10k dataset. The evaluation split consisted of 100 lesions requiring clinical intervention (30 melanomas, 25 basal cell carcinomas, 25 squamous cell carcinomas, 20 actinic keratoses) and 100 safe to discharge (50 benign naevi, 50 seborrhoeic keratoses).
The baseline model, Google's 2015 Inception v3, was fine-tuned on 68,011 clinical and dermoscopic images from the ISIC Archive on a single NVIDIA H100 GPU for 20 minutes. Test lesions and patient matches were strictly excluded from training data.
Accuracy on 200 Biopsy-Verified Lesions:
- Fine-tuned Inception v3: 125 / 200 (62.5%)
- Claude Fable 5.1: 85 / 200 (42.5%)
- GPT-6 Astra: 84 / 200 (42.0%)
Second-Choice Diagnostic Accuracy:
- Fine-tuned Inception v3: 173 / 200 (86.5%)
- Claude Fable 5.1: 138 / 200 (69.0%)
- GPT-6 Astra: 134 / 200 (67.0%)Systematic Failure Modes in Frontier Multimodal Models
Frontier vision-language models suffered from extreme label bias. GPT-6 Astra classified 75 lesions as basal cell carcinoma when only 25 existed in the sample, mistakenly including 14 of the 30 melanomas. Claude Fable 5.1 labeled 53 lesions as actinic keratosis against an actual prevalence of 20.
On harmless lesions, the CNN accurately identified 37 of 50 moles and 26 of 50 seborrhoeic keratoses, whereas GPT-6 Astra recognized 16 and 13 respectively.
Deployment Architecture: Edge vs API Latency
Unlike multimodal LLMs requiring remote API requests and recurring token charges, the fine-tuned convolutional model compiles to run client-side in a mobile browser using WebGPU. For production classification pipelines where spatial inductive biases matter, domain-adapted specialized encoders consistently beat generalized frontier models while eliminating network hops, payload serialization, and privacy risks.
✓ When to use
- Use fine-tuned convolutional networks for closed-set visual classification where latency, privacy, and inference compute budgets are constrained.
- Deploy lightweight vision encoders when inference must execute client-side via WebGPU on edge hardware or mobile browsers.
✕ When NOT to use
- Do not use small CNNs when open-ended visual reasoning, natural language report synthesis, or unconstrained free-text conversation is required.
- Avoid specialized classification models when labeled domain training datasets are unavailable or prohibitively costly to acquire.
What to do today
- Audit existing vision workflows to identify classification tasks currently routed to costly multimodal LLM APIs.
- Benchmark transfer learning on lightweight architectures like MobileNet or Inception v3 before provisioning frontier vision tokens.
- Evaluate client-side WebGPU runtime engines for latency-sensitive and privacy-critical image inference pipelines.
What the community says
“Training a model requires labels training data. Getting up and running with an LLM just requires a prompt. I agree that a good model will outperform an LLM, but the LLM is so much faster/easier to build.”
“Internet discovers If statements subnanosecond response times model fits on CPU”
“The tooling around LLMs is still nascent and, in my opinion, quite poor... an LLM plus tooling dedicated to the task plus a human in the loop will open up capabilities not available to an LLM in a chat.”