A 10-Year-old CNN Outperforms Frontier LLMs For Skin Cancer Diagnosis

TL;DR

Before you orderOffer from Amazon

Get workout gear delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

A dermatologist reports that a fine-tuned, 10-year-old Inception v3 convolutional neural network correctly identified 125 of 200 biopsy-confirmed skin lesions, compared with 84 and 85 for two large language models. The test was small and used a selected public dataset, so its results do not establish how the systems would perform in routine care.

A dermatologist reports that a fine-tuned Inception v3 convolutional neural network, originally published in 2015, correctly named 125 of 200 biopsy-confirmed skin lesions in a test, outperforming two tested frontier large language models, which scored 84 and 85. The comparison, published on Lynch Lab’s website, is a limited evaluation of selected images—not evidence that the older system is ready to diagnose patients or that the results apply to clinical practice.

The dermatologist tested GPT-6 Astra and Claude Fable 5.1 alongside a version of Inception v3 fine-tuned on dermatology images. Each case included a clinical photograph and a dermoscopic image, taken through the magnifying lens used by dermatologists. The LLMs were asked to choose among six diagnoses or say they were unsure, give confidence and a second choice, and recommend discharge, biopsy or follow-up. The report says both models received the same prompt at default settings and neither declined a case.

The 200 lesions were drawn at random from MILK10k, a public dataset released through the International Skin Imaging Collaboration in 2025. All had histopathology results. The sample comprised 100 lesions the author classified as not safe to discharge—30 melanomas, 25 basal cell carcinomas, 25 squamous cell carcinomas and 20 actinic keratoses—and 100 lesions considered safe to discharge: 50 naevi and 50 seborrhoeic keratoses. All had been biopsied, including the benign lesions.

Overall, Inception v3 named the correct diagnosis in 62% of cases; GPT-6 Astra did so in 42% and Claude Fable 5.1 in 42.5% (85 of 200, rounded to 42% in the report). On the melanoma subset, the CNN was correct on 13 of 30, compared with 8 of 30 for GPT-6 Astra and 11 of 30 for Claude. In a highlighted example, both LLMs labeled an invasive melanoma on the ear as a harmless seborrhoeic keratosis and said it could be discharged; the CNN identified melanoma from the dermoscopic image.

A Reminder That Image Tasks Differ

The result challenges the assumption that a newer, general-purpose language model will automatically outperform an older, specialized vision system on a medical image task. In this comparison, the CNN’s overall score was higher, but performance varied by diagnosis: GPT-6 Astra correctly classified all 25 basal cell carcinomas in the sample, while the CNN correctly classified 24. The findings do not support a simple ranking that would hold across every condition.

For readers, the practical point is that a high-profile model’s fluency or confidence is not a substitute for validated diagnostic performance. The report’s melanoma example also illustrates the stakes of an incorrect reassurance: a system that suggests a lesion is harmless could influence whether a person seeks assessment. No model in this test should be treated as a replacement for a clinician, and the results do not establish that any of the systems can safely guide discharge decisions.

The author also says the CNN was small enough to run in a phone browser, with images processed locally rather than uploaded. That may be relevant to privacy and access, but a demonstration of local operation does not establish clinical safety, reliability across devices, or suitability for patient use.

How the Three Systems Were Compared

The CNN was Inception v3, a network first published by Google in 2015 and originally pretrained on everyday photographs. According to the report, the author fine-tuned all its layers for about 20 minutes on a single H100 GPU using 68,011 clinical and dermoscopic images from the ISIC Archive. The 200 test lesions, as well as other photographs of the same lesions or patients, were excluded from training and validation, the author says.

The network assessed both images for each lesion, placing most weight on the dermoscopic image, and chose among the same six diagnoses. The comparison therefore tested a tailored image-classification model against two general-purpose LLMs given paired images and a clinical-style prompt. It did not compare the systems in a live clinic, with patient histories, physical examinations or clinician oversight. The test set was deliberately composed of biopsied lesions, which the author notes can be harder to diagnose than lesions not selected for biopsy.

Limits of the 200-Lesion Test

The report describes a single evaluation of 200 lesions, not a large, independent clinical trial. It does not establish how the models would perform on images from different cameras, clinics, skin tones, or patient populations, nor whether results would hold when clinicians provide medical history and other information. The source material does not report confidence intervals or a formal statistical comparison, so the precision of the difference between these scores is unclear.

Although the lesions were randomly drawn from the stated dataset, all were biopsied. That selection means the sample may not reflect the range or prevalence of lesions seen in routine appointments. The report also does not provide enough information to assess calibration of the systems’ confidence estimates or how often their discharge recommendations would have led to missed disease across a representative patient population. Independent replication and prospective clinical testing are not reported.

Independent Testing Before Patient Use

The report does not announce a clinical trial, regulatory review or planned deployment. The immediate next step for evaluating the claim would be independent replication on larger, diverse datasets, followed by testing under clinical conditions where image quality, patient histories and clinician decisions can be accounted for. Such work would need to assess not only the correct diagnosis but also errors that could lead to inappropriate reassurance or unnecessary procedures.

The author says readers can try the CNN through a browser-based demonstration that draws from the same 200 lesions. That offers a way to inspect the reported test, but it is not a validation study or a patient-care tool. Until additional evidence is available, people concerned about a skin lesion should seek assessment from a qualified health professional rather than rely on an AI result.

Key Questions

Which system scored highest in the reported test?

The fine-tuned Inception v3 CNN correctly named 125 of 200 lesions, or 62%. GPT-6 Astra scored 84 of 200, while Claude Fable 5.1 scored 85 of 200.

Does this show that a 10-year-old CNN is better than all frontier LLMs?

No. It compares one CNN with two named LLMs on one set of 200 lesions. The results do not establish performance for other models, datasets or clinical settings.

What kinds of skin lesions were included?

The test included six diagnoses: melanoma, basal cell carcinoma, squamous cell carcinoma, actinic keratosis, naevi and seborrhoeic keratoses. Every lesion in the sample had been biopsied and confirmed by histopathology.

Can patients use the CNN to decide whether a lesion needs care?

The report does not establish that the CNN is safe for diagnosis or discharge decisions. Its browser demonstration uses the same test cases, and does not replace assessment by a qualified health professional.

Source: hn