An educational machine learning demonstration framework for plastic surgeons using open datasets.

[BACKGROUND] Artificial intelligence (AI) offers powerful tools for dermatology and plastic surgery, but practical engagement with data science remains limited. Open datasets now allow clinicians to explore machine learning without specialist infrastructure. This study presents DermAI-Melanoma, an open data demonstration framework for plastic surgeons that uses melanoma classification as a didactic case study to demonstrate reproducible training and deployment of deep learning models using a public dataset.

[METHODS] The SIIM-ISIC 2020 melanoma dataset was used. To avoid data leakage, patient-level stratification was applied so that all images from 1 patient were assigned to either training or testing. Images underwent resizing, color balancing, and augmentation. Two convolutional neural networks were trained: (1) EfficientNet-B3 benchmark model and (2) a lightweight MobileNetV3-Small optimized for smartphones. Both models were converted to TensorFlow.js for browser-based deployment.

[RESULTS] The EfficientNet-B3 achieved 97% test accuracy, detecting 92% of melanomas and correctly classifying 96% of benign lesions (area under the receiver operating characteristic curve [AUC-ROC] = 0.98, F1-score = 0.91). The MobileNetV3-Small achieved 94% accuracy and required <5 MB storage, producing results in under 2 s on a standard smartphone. Model performance was comparable to dermatologist benchmarks in literature on melanoma detection tasks in published studies (range: 65-79% with visual inspection alone, increasing to 82-91% with dermoscopy).

[CONCLUSION] DermAI-Melanoma demonstrates how plastic surgeons can actively engage with data science by using open datasets to build transparent, deployable AI tools. Beyond melanoma, similar frameworks could support education, research, and innovation in plastic surgery, provided the specialty embraces the importance of open data sharing and cross-disciplinary collaboration.