ISIC-2024 SLICE-3D

Open the ISIC-2024 SLICE-3D site

A single-dataset, no-external-data, no-synthetic study of melanoma triage from 3D total-body photography. The task is to classify 393 malignant lesions among 400,666 benign crops using only the ISIC-2024 SLICE-3D dataset, with no external dermoscopy archives and no diffusion-synthesised positives.

The ISIC-2024 leaderboard winners reached pAUC@80%TPR ≈ 0.173, but only by importing external dermoscopy data and roughly 30,000 synthetic malignant lesions. This work bans both, and asks a more specific question: how much pAUC does each unit of inference cost yield on SLICE-3D alone? The project reports a quality vs. efficiency frontier rather than a single number.

Headline result: 0.1738 stack OOF pAUC@80%TPR at 15.84M parameters, 2.46 GFLOPs, and 60.9 ms CPU latency per inference. The tabular GBDT baseline alone scores 0.1689 and the image expert alone scores 0.1582. The pipeline combines a LightGBM tabular expert over intrinsic engineered features (geometry, L*a*b* colour, border and shape, 3D body position, patient-relative “ugly-duckling” deviation, fold-local target encoding) with a small ImageNet-pretrained ConvNeXt-V2-nano image backbone at 224 px. The image expert is trained with AdamW and cosine schedule under a heavy augmentation stack (brightness/contrast, blur/noise, distortions, CLAHE, hue-saturation, shift-scale-rotate, coarse-dropout) alongside mixup α=0.2, label smoothing 0.05, weight EMA 0.995, and negative undersampling. A trivial rank-average combiner produces the final score.

Validation uses patient-grouped, target-stratified 5-fold cross-validation frozen at initialization, with the official pAUC scorer independently verified as numerically identical to the implementation. The README also documents negative results: learned mixture-of-experts gating, meta-learner stacking, embedding injection, heavy transformers, and aggressive EMA smoothing all degraded performance against simpler alternatives.

It does not beat the unconstrained winner’s private pAUC, but within the no-external-data, no-synthetic, single-dataset setting it is the best reported result.

Produced as a Neural Networks course project at National Yunlin University of Science and Technology by team Pakistan.AI: Muhammad Junaid Ali Asif Raja, Adil Sultan, and Shahzaib Ahmed Hassan. Instructor: Prof. Hsuan-Ting Chang. This is applied efficient-deep-learning work, separate from the fractional machine learning that anchors the rest of this site.

Companion website: junaidaliop.github.io/isic2024-tbp
Source code: github.com/junaidaliop/isic2024-tbp