At the University of Missouri in Columbia, a research team says an artificial intelligence ensemble can flag melanoma in photos with roughly 92 percent accuracy after training on hundreds of thousands of skin images. The researchers combined three different neural network architectures into one model and tested it on 3D total body photography, where the ensemble outperformed the individual models. They describe the system as a decision support tool meant to speed evaluation of suspicious lesions rather than a replacement for clinicians.
Kamlendra Singh, an associate research professor who led the project, has called the system a promising “proof of concept” and says it is aimed at improving access to dermatology care. The team trained and tested its models on a database of about 400,000 skin abnormality images captured with 3D total body photography, according to Show Me Mizzou. Singh has said the next steps include making the tool more explainable and validating performance across more diverse skin tones and camera conditions.
Study details and technical performance
The peer reviewed paper, published in Biosensors and Bioelectronics: X, reports that an ensemble built from ConvNeXt, ResNet 50 and Swin Transformer architectures achieved an area under the curve (AUC) of 0.9208 in cross validation. The journal lists individual model AUCs of roughly 0.85 to 0.88, which the authors say highlights the benefit of ensemble learning for this task. The paper also notes that the team used publicly available 3D total body photography data and a leak free cross validation approach to limit overfitting.
What researchers say and real world limits
Singh has emphasized that “the goal is not for AI to replace doctors” and that the project is intended to support clinicians by speeding triage and referral decisions, according to Show Me Mizzou. Independent coverage has summarized similar caveats about how well the system will generalize and the need for clinical validation, per MedicalXpress. Dermatologists and regulators typically expect prospective clinical testing and close checks for bias before research models are turned into patient facing tools.
Next steps and a consumer app on the radar
Local reporting says the MU team is working toward a phone app that could let users snap a picture for an initial screen, though researchers stress that any consumer facing product would require extensive testing and oversight, as reported by KOMU. The underlying study is available in the journal (DOI 10.1016/j.biosx.2025.100714), which lists an online publication in late 2025, and the university ran a January 2026 press feature summarizing the work. For now, the team presents the results as an encouraging step toward faster detection and better triage of suspicious lesions.
