# Is Dataset Quality Still a Concern in Diagnosis Using Large Foundation Model?

**Authors:** Ziqin Lin, Heng Li, Zinan Li, Huazhu Fu, Jiang Liu  
**arXiv:** [2405.12584](https://arxiv.org/abs/2405.12584)  
**Date:** 2024-05-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2405.12584v1/

## Key points

- Study title: "Is Dataset Quality Still a Concern in Diagnosis Using Large Foundation Model?"
- Authors: Ziqin Lin, Heng Li, Zinan Li, Huazhu Fu, Jiang Liu
- Focus on impact of dataset quality on performance of pre-trained large foundation models (LFM) in medical diagnostic tasks
- Development of LFM using Vision Transformer (VIT) and self-supervised learning framework for fundus image analysis
- Comparison of LFMs with traditional convolutional networks in handling dataset quality issues
- Finding that LFMs are more resilient to dataset quality issues
- Effectiveness of overall fine-tuning strategy to adapt LFMs and reduce impact of dataset quality challenges
- Key questions addressed: Resilience to image quality variations, effect of dataset bias on LFM performance, role of fine-tuning techniques in mitigating effects
- Potential of LFMs in handling data variability emphasized

## Summary

In their study titled "Is Dataset Quality Still a Concern in Diagnosis Using Large Foundation Model? ", authors Ziqin Lin, Heng Li, Zinan Li, Huazhu Fu, and Jiang Liu explore the impact of dataset quality on the performance of pre-trained large foundation models (LFM) in medical diagnostic tasks. The researchers highlight recent advancements in LFMs and their significant breakthroughs in natural language processing and computer vision domains. Focusing specifically on fundus image analysis for disease diagnosis, the authors developed an LFM using the Vision Transformer (VIT) and a self-supervised learning framework. The study investigates how LFMs handle dataset quality issues such as image quality and dataset bias by conducting experiments with varying quality datasets in fundus diagnosis tasks. Through their investigation, the authors found that LFMs demonstrate greater resilience to dataset quality issues compared to traditional convolutional networks. They also discovered that overall fine-tuning is an effective strategy to adapt LFMs and reduce the impact of dataset quality challenges. The research aims to answer key questions: Is LFM more resilient to image quality variations? How does dataset bias affect LFM performance? Can fine-tuning techniques help mitigate these effects? The study sheds light on the potential of LFMs in handling data variability and highlights the importance of considering dataset quality when developing models for medical diagnostics using large foundation models.