Diagnostic performance and energy consumption of deep learning models for dermoscopic image classification: data, code and results

Replication package for a study comparing ResNet-50, EfficientNet-B3 and Vision Transformer B/16 on dermoscopic image classification (HAM10000), evaluating diagnostic performance, calibration, and directly measured energy consumption at both training and inference time. Two methodological features distinguish this analysis. The dataset is partitioned at the lesion level, grouped on lesion_id, so that no lesion appears in more than one partition; HAM10000 contains 10,015 images derived from 7,470 unique lesions, and an image-level split allows different images of the same lesion into both training and test sets. GPU energy is measured directly by sampling board-level power through the NVIDIA management interface at 100 ms intervals and integrating over the run, rather than estimated from software telemetry or thermal design power. Contents: predicted probabilities on the held-out test set for three architectures across three independent training runs each (nine files); the lesion-level test split definition; per-run and ensembled result files; seven result tables in CSV; six figures; and the full reproduction pipeline. The deposited predictions are sufficient to reproduce every reported performance figure without retraining and without access to the source images. Source images from HAM10000 (CC BY-NC 4.0) are not redistributed. Code is licensed MIT; derived data, results and figures are CC BY 4.0.