Abstract

Long-tailed class distributions are a defining challenge in medical image classification, especially in resource-constrained settings where limited data collection thins the tail of rare but clinically important classes. Diffusion-based synthetic augmentation is promising, but existing approaches rely on costly generative training and on a single, manually chosen OOD-filtering fraction. We study a parameter-efficient pipeline for HAM10000 skin lesion classification that adapts a frozen inpainting diffusion backbone with per-class Low-Rank Adaptation (LoRA), automatically selects an out-of-distribution (OOD) post-processor from a pool of established methods, and treats the clean-synthetic fraction $\gamma$ as a tunable operating point. In a single-split HAM10000 experiment, automatic selection identifies Mahalanobis scoring as the highest-AUROC filter (0.994), ahead of KNN (0.943), ViM (0.856), MSP (0.632), ODIN (0.610) and Energy (0.493). A sweep over $\gamma \in {0.2, 0.4, 0.6, 0.8, 1.0}$ affects balanced multiclass accuracy (BMA) non-monotonically: the highest observed value is 0.848 at $\gamma=0.6$, against 0.836 for the real-only baseline and 0.832 at $\gamma=1.0$. There, recall exceeds baseline for vascular lesions ($+10.7$ points), actinic keratoses ($+1.5$) and benign keratosis-like lesions ($+0.9$), but melanoma recall is \emph{below} baseline at every $\gamma$ tested (smallest gap $-0.4$, one image, at $\gamma=0.8$). Since every configuration was trained once on one split, and $\gamma$ was chosen on the fold where these numbers are measured, the $1.2$-point BMA difference is an observation, not a statistically established improvement; the split is also image-level rather than lesion-grouped ($38.7\%$ of evaluation images share a lesion with training), so absolute values are optimistic. The class-level pattern is the more robust finding: synthetic augmentation reallocates sensitivity across classes in a $\gamma$-dependent way, and a $\gamma$ chosen by aggregate accuracy alone can carry a cost on the most clinically consequential class.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/MIRASOL_062.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/MIRASOL_062_supp.pdf

Link to Open Review

Open Review Page: https://openreview.net/forum?id=JIRzVrc5HM

BibTex

@InProceedings{GalIgn_Efficient_MICCAISAT2026,
        author = { Gallo, Ignazio AND Sgroi, Leonardo AND Gatti, Mattia},
        title = { { Efficient Diffusion-Based Synthetic Augmentation with Automatic OOD-Driven Filtering for Long-Tailed Skin Lesion Classification in Resource-Constrained Settings } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 17264},
        month = {pending},
        page = {pending}
}


back to top