Related Experiment Video
Updated: May 12, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.6K
Effective data selection via deep learning processes and corresponding learning strategies in ultrasound image
Hyunju Lee1, Jin Young Kwak1, Eunjung Lee2
1Department of Radiology, Severance Hospital, Research Institute of Radiological Science, College of Medicine, Yonsei University, 50 Yonsei-ro, Seodaemun-gu, Seoul, 03722, Republic of Korea.
Scientific Reports
|May 9, 2025
Summary
This study introduces a novel transfer learning method for medical image classification, optimizing data selection with deep learning. It enhances performance in data-limited scenarios by using a "True network" to refine initial classifications, improving accuracy without new data.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Computer Vision
Background:
- Transfer learning is crucial for medical image classification, but performance often plateaus with limited datasets.
- Existing methods struggle to maximize learning from constrained datasets, hindering progress in data-scarce medical applications.
- Optimizing data selection within deep learning frameworks is key to overcoming data limitations in medical AI.
Purpose of the Study:
- To develop and evaluate a novel approach for enhancing transfer learning in medical image classification by optimizing data selection.
- To address the challenge of limited datasets by maximizing the utility of existing data through innovative deep learning strategies.
- To improve classification accuracy, sensitivity, specificity, F1-score, and AUC in medical imaging tasks without requiring additional data.
Main Methods:
- A two-stage approach involving an original classification network and a secondary 'True network'.
- The True network, sharing the original network's architecture, is trained on a consensus-selected data subset to verify and correct classifications.
- Experiments utilized ResNet101, Vision Transformer, and other pre-trained networks on thyroid nodule ultrasound and dermoscopic images.
Main Results:
- The proposed method significantly improved all five key performance metrics (accuracy, sensitivity, specificity, F1-score, AUC) on ResNet101 compared to baseline networks.
- Similar performance enhancements were observed when applying the True network to Vision Transformer and various convolutional neural network architectures.
- The approach demonstrated robustness and adaptability across different medical imaging modalities, including ultrasound and dermoscopic images.
Conclusions:
- The novel data selection and dual-network strategy effectively enhances transfer learning for medical image classification in data-limited settings.
- This method provides a viable solution for improving diagnostic accuracy without the need for acquiring new or expanded datasets.
- The findings highlight the potential of intelligent data optimization for advancing AI in medical diagnostics.
