Experimental Data and Source Code for Oracle-Free Metamorphic Test Case Prioritization for Deep Learning-Based Medical Image Classification

This code and supplementary-material package supports the experiments reported in the manuscript prepared for the Journal of Software: Evolution and Process. The code implements the oracle-free metamorphic test case prioritization workflow for CIFAR-10 and ISIC2019. It includes the proposed UI-SumRank method, which combines uncertainty variation based on Jensen-Shannon divergence and interpretability variation based on Grad-CAM drift, as well as the ablation variants U-only, I-only, and equal-weight UI. The baseline methods DeepGini, QuoTe, and Random are also included. The code supports metamorphic-relation image generation, class-balanced Top-K prioritization, seed-level experiment execution, runtime profiling, and post-hoc statistical analysis. The main scripts include generate_cifar10_isic_mrs.py for generating metamorphic image variants; ui_sumrank_cifar10.py and ui_sumrank_isic2019.py for the proposed method and ablations; cifar10_baselines_all.py and isic2019_baselines_all.py for baseline experiments; batch_run_main.py and batch_run_all_baselines.py for batch execution; runtime_utils.py for runtime logging; and analyze_runtime_and_statistics.py for aggregating runtime, mean, standard deviation, confidence interval, and effect-size results. The supplementary archive JSEP_supplementary_individual_pdfs.zip contains individual PDF files for the supplementary figures and tables associated with the manuscript. It includes Figure S1, which reports Top-50 prioritization result plots; Figure S2, which presents the critical difference diagrams used for non-parametric comparison; Figure S3, which reports the Top-50 ablation plots; and Table S1, which provides the runtime breakdown. These files are provided separately to support journal submission, reviewer inspection, and reproducibility of the reported experimental analysis. Raw CIFAR-10 and ISIC2019 images and trained model checkpoints are not redistributed because they are third-party resources or locally trained artifacts. Users should download or prepare the datasets and checkpoints locally, then use the provided scripts to reproduce the ranking records, seed-level summaries, runtime profiles, aggregated statistical outputs, and supplementary figures/tables.本代码及辅助材料包可复现提交至《软件：演化与过程》（Journal of Software: Evolution and Process）的论文中所报道的各项实验。本代码实现了针对CIFAR-10与ISIC2019数据集的无预言蜕变测试用例优先级排序工作流，其中包含本文提出的UI-SumRank方法——该方法结合了基于詹森-香农散度（Jensen-Shannon divergence）的不确定性变化量与基于Grad-CAM漂移的可解释性变化量，同时还提供了其消融变体：仅不确定性（U-only）、仅可解释性（I-only）以及等权重UI（equal-weight UI）。此外，代码还涵盖了DeepGini、QuoTe与Random这三种基准方法。 本代码支持元蜕变关系图像生成、类别平衡Top-K优先级排序、种子级实验执行、运行时性能剖析以及事后统计分析。核心脚本包括：用于生成元蜕变图像变体的generate_cifar10_isic_mrs.py；用于实现本文提出方法与消融实验的ui_sumrank_cifar10.py与ui_sumrank_isic2019.py；用于基准实验的cifar10_baselines_all.py与isic2019_baselines_all.py；用于批量执行任务的batch_run_main.py与batch_run_all_baselines.py；用于运行时日志记录的runtime_utils.py；以及用于汇总运行时数据、均值、标准差、置信区间与效应量结果的analyze_runtime_and_statistics.py。 辅助归档文件JSEP_supplementary_individual_pdfs.zip包含与该论文相关的补充图表PDF文件，具体包括：展示Top-50优先级排序结果可视化图的图S1；用于非参数比较的临界差异图S2；展示Top-50消融实验可视化结果的图S3；以及提供运行时细分数据的表S1。本辅助文件单独提供，以支持期刊投稿、审稿人审阅以及所报道实验分析的可复现性。 原始CIFAR-10与ISIC2019图像以及预训练模型检查点不属于本包分发范围，因其属于第三方资源或本地训练得到的成果。用户需自行本地下载或准备对应数据集与模型检查点，随后通过本包提供的脚本即可复现优先级排序结果、种子级实验汇总、运行时性能剖析数据、汇总统计输出以及补充图表。