# FICAug: Feature-Informed Clustering and Augmentation for Facial-Expression-Based Parkinson's Disease Screening

arXiv: https://arxiv.org/abs/2409.17685 · HTML: https://mlsanity.com/paper/2409.17685/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 9 of its 35 references that are in the map.
- First posted on arXiv 26 Sep 2024.

## Builds on (most cited first, 9 of 9)

- [Deep Residual Learning for Image Recognition](https://mlsanity.com/paper/1512.03385/) (2015)
- [Emerging Properties in Self-Supervised Vision Transformers](https://mlsanity.com/paper/2104.14294/) (2021)
- [Improved Regularization of Convolutional Neural Networks with Cutout](https://mlsanity.com/paper/1708.04552/) (2017)
- [Masked Siamese Networks for Label-Efficient Learning](https://mlsanity.com/paper/2204.07141/) (2022)
- [StableRep: Synthetic Images from Text-to-Image Models Make Strong Visual Representation Learners](https://mlsanity.com/paper/2306.00984/) (2023)
- [GANimation: Anatomically-aware Facial Animation from a Single Image](https://mlsanity.com/paper/1807.09251/) (2018)
- [Learning Vision from Models Rivals Learning Vision from Data](https://mlsanity.com/paper/2312.17742/) (2023)
- [Deep Convolutional Neural Network with Mixup for Environmental Sound Classification](https://mlsanity.com/paper/1808.08405/) (2018)
- [From Fake to Real: Pretraining on Balanced Synthetic Images to Prevent Spurious Correlations in Image Recognition](https://mlsanity.com/paper/2308.04553/) (2023)

## Abstract (arXiv)

Hypomimia has drawn growing interest as a digital marker for screening Parkinson's disease (PD). However, developing reliable facial-expression-based screening models is challenging because clinical PD datasets are small, exposing models to only a narrow range of how hypomimia can appear across individuals. Standard augmentation strategies do not solve this problem; recombining or perturbing existing samples produces variation, but not new facial configurations that are plausible and clinically meaningful. We introduce FICAug to address this gap. The framework clusters Action Unit (AU) feature vectors extracted from facial expression images, discards clusters that mix labels inconsistently, and generates synthetic AU vectors within the retained clusters with Gaussian sampling. GANimation then reconstructs these synthetic vectors into realistic facial images. A ResNet18 model is pretrained on these reconstructed images, and then fine-tuned on real clinical data. Using the UT-MoDaPark dataset, FICAug achieved 88.63% cross-validation accuracy and 94.00% test accuracy, outperforming both a standard ResNet18 baseline and self-supervised alternatives, including DINO and MSN. These results suggest that synthetic image generation, when guided by label consistency and clinical feature structure, can function as an effective intermediate representation-learning step for facial-expression-based PD screening in settings where clinical data remain scarce.