Combined Oriented Data Augmentation Method for Brain MRI Images
In recent years, deep learning’s use in medical imaging has grown exponentially. However, one of the biggest problems with training deep learning models is the unavailability of large amounts of data, which leads to overfitting. Collecting large quantities of labelled medical images is ex...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10829922/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In recent years, deep learning’s use in medical imaging has grown exponentially. However, one of the biggest problems with training deep learning models is the unavailability of large amounts of data, which leads to overfitting. Collecting large quantities of labelled medical images is expensive, time-consuming, and depends on specialists’ availability. In this paper, we proposed a novel method namely Oriented Combination MRI (OCMRI) for augmenting brain MRI dataset. The proposed method helps CNN models overcome overfitting and address class imbalance issues by combining Brain MRI images to generate new images. The image fusion is performed by selecting two images of the same tumor class if the Mean Squared Error (MSE) between these two images is greater than threshold 1 and lower than threshold 2. Both thresholds are adjustable, initially set by the user and automatically fine-tuned by the algorithm to control the number of images produced for each class, thus helping to address the data imbalance problem. The proposed approach was evaluated by training and testing the PRCnet model on four publicly available datasets before and after applying the proposed method to the datasets. Where the classification accuracy without data augmentation was 85.19% for dataset A, 90.12% for dataset B, 94.77% for dataset C, and 90% for dataset D respectively. After adding the synthetic data; the accuracy improved to 92.7% for dataset A, 95.37% for dataset B, 96.51% for dataset C and 98% for dataset D respectively. |
|---|---|
| ISSN: | 2169-3536 |