Skin-Lesion Classification

Skin-Lesion Classification Supporting material for Group-Aware Evaluation and Experiment-Provenance Auditing for Seven-Class Skin Lesion Classification on HAM10000. The study audits 14 historical runs, examining evaluation integrity, class imbalance, and whether reported metrics belong to the same epoch. Supplied files File Contents skin-lesion-code-audit-v1.0.0.zip Historical training scripts, P2 source-code export, implementation audit, and reporting-verification tools. skin-lesion-results-provenance-v1.0.0.zip Training-log evidence, configurations, 150 P2 epoch records, 20 paired metric selections, and reproduced tables. manuscript-revision.zip Revised LaTeX manuscript, references, replacement sections, correction notes, and an eight-page compiled draft. experimental_section.tex Replacement experimental-configurations section; identical to the copy inside the manuscript ZIP. Towards_Robust_Skin_Lesion_Classification (7).pdf Supplied seven-page paper; a different draft from the PDF inside the manuscript ZIP. Main files inside the archives File or folder Purpose main.tex, references.tex, main.pdf Complete revised manuscript, bibliography, and compiled PDF. experimental_section.tex, results_section.tex Replacement sections for configurations and results. URGENT_FIXES.md Explains manuscript corrections. historical_repository/train*.py Preserved training variants using different losses, sampling, and evaluation settings. historical_repository/threshold_calibration.py Selects a melanoma decision threshold using validation results. HISTORICAL_SOURCE_AUDIT.md Documents implementation differences and reproducibility issues. evidence/ Historical summaries, configurations, epoch histories, and selected metric pairs. Use p2_selected_epochs.csv for paired results. source/, work/ Preserved source PDFs and extracted text. audit/reproduce_tables.py, generated/ Reporting-check script and its generated tables and summaries. verify_checksums.py, MANIFEST.sha256 File-integrity verification. The code and results archives intentionally share audit scripts and evidence tables. Usage For Overleaf, upload manuscript-revision.zip and select main.tex. The separate section files require the manuscript context. Inside either extracted code/results package, run with Python 3.9+: python3 verify_checksums.py python3 audit/reproduce_tables.py --output-dir reproduced