UCDSC: Open Set UnCertainty aware Deep Simplex Classifier for Medical Image Datasets

This repository contains the implementation code for the paper "UCDSC: Open Set UnCertainty aware Deep Simplex Classifier for Medical Image Datasets". Link to arxiv preprint: https://arxiv.org/abs/2511.08196

Dataset Setup

MedMNIST Dataset

The MedMNIST dataset can be downloaded from:

• Official source: https://zenodo.org/records/10519652

Background Data

For extended experiments, download the 300K random images:

• Source: https://people.eecs.berkeley.edu/~hendrycks/300K_random_images.npy

Augmented Skin Conditions dataset

• Source: https://www.kaggle.com/datasets/syedalinaqvi/augmented-skin-conditions-image-dataset/data

Basic Training

Run the main training script with default parameters:

python NirvanaOSR.py --dataset dataset-name --dataroot ./data --outf ./results

Hyperparameter Configuration

Parameter	Description
--batch-size	Training batch size
--lr	Learning rate
--max-epoch	Maximum training epochs
--optim	Optimizer to be used
--margin	Margin for loss
--Expand	Expand factor of centers
--uncertainty-weight	Weight for uncertainty loss
--outlier-weight	Weight for outlier triplet loss
--model	Backbone network to be used

Acknowledgments

This codebase is derived from the Deep Simplex Classifier implementation available at https://github.com/Cevikalp/dsc. We thank the authors for making their code available.