Diagnosis of cerebral microbleed via VGG and extreme learning machine trained by Gaussian map bat algorithm

Siyuan Lu, Kaijian Xia, Shui-Hua WangJ. Ambient Intell. Humaniz. Comput 2023 · OpenReview

TL;DR

This paper introduces a novel deep learning approach for automatically diagnosing cerebral microbleeds (CMBs) from MRI scans. By combining a pre-trained VGG network for feature extraction with an Extreme Learning Machine (ELM) optimized by a Gaussian Map Bat Algorithm (GBA), the method achieves high accuracy and sensitivity, outperforming existing state-of-the-art techniques.

What Problem Does This Paper Address?

Cerebral microbleeds (CMBs) are small lesions in the brain associated with serious health conditions like dementia. Detecting them accurately from MRI scans is crucial for diagnosis and treatment. However, manual inspection by radiologists is a time-consuming and subjective process, leading to potential inter-observer variability and missed diagnoses. Existing automated methods often struggle with reproducibility or require extensive datasets for training deep learning models, risking overfitting. This paper aims to develop a robust and efficient automated system for CMB diagnosis that overcomes these limitations.

Methods

The proposed method, dubbed VGG-ELM-GBA, integrates several key components:

  • Feature Extraction with Pre-trained VGG: Instead of training a deep convolutional neural network (CNN) from scratch on a potentially small CMB dataset, the authors leverage a pre-trained VGG-16 network. This network, trained on the large ImageNet dataset, is used to extract rich image features from the MRI scans. The last two layers of VGG are removed to obtain these features.
  • Extreme Learning Machine (ELM) for Classification: For the classification task, an ELM is employed. ELMs are known for their rapid training speed (significantly faster than traditional backpropagation algorithms) and good generalization performance, making them suitable for scenarios with limited data.
  • Gaussian Map Bat Algorithm (GBA) for Optimization: To further enhance the ELM's performance and robustness, the hidden weights and biases of the ELM are optimized using a Bat Algorithm (BA) that incorporates a Gaussian chaotic map. This chaotic map helps the BA explore the solution space more effectively, preventing it from getting stuck in local optima and leading to better parameter tuning. The paper explores three variants of the Bat Algorithm (Gaussian, Logistic, and Cubic maps) and finds the Gaussian map to be optimal for their task.
  • Sliding Neighborhood Processing (SNP): To generate training and testing samples from the MRI volumes, a sliding window approach is used. This method creates smaller image patches, and their labels are determined by the central voxel's classification.

Datasets

The study utilized a dataset comprising:

  • Ten CADASIL (Cerebral Autosomal Dominant Arteriopathy with Subcortical Infarcts and Leukoencephalopathy) samples.
  • Ten healthy control samples.

These samples were reconstructed and labeled by experienced medical experts. The dataset was pre-processed using sliding neighborhood processing to generate a total of 13,031 samples, with 6407 labeled as CMB and 6624 as non-CMB. 70% of this dataset was used for training, and the remaining 30% for testing.

Key Insights and Ideas

  • Leveraging Pre-trained Models for Limited Data: The successful use of a pre-trained VGG network highlights the power of transfer learning. It allows for effective feature extraction even when the target dataset is relatively small, mitigating the risk of overfitting that often plagues deep learning models trained from scratch on limited data.
  • Synergy of Feature Learning and Optimization: The combination of a powerful feature extractor (VGG) with a fast and effective classifier (ELM) that is then meticulously optimized by a metaheuristic algorithm (GBA) proves to be a potent strategy. This hybrid approach captures the strengths of both deep learning for representation learning and swarm intelligence for parameter optimization.
  • The Role of Chaotic Maps in Optimization: The integration of chaotic maps into the Bat Algorithm demonstrates a sophisticated approach to enhancing optimization. The Gaussian map, in particular, was found to improve the algorithm's ability to escape local minima and find a more globally optimal solution for ELM parameters.

Key Results

The proposed VGG-ELM-GBA method achieved impressive results:

  • Sensitivity: 93.08%
  • Specificity: 87.12%
  • Accuracy: 90.05%

These metrics indicate a strong ability to correctly identify CMBs and non-CMBs. The paper also shows that the 'fc8' layer of VGG provided the optimal features, and the Gaussian map was the best performing chaotic map for optimizing the ELM. The method's convergence time was reported as 170.57 seconds, which is considered affordable for practical applications.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Interpretability: The "black box" nature of deep learning models makes it difficult to interpret how the diagnosis is made.
  • Binary Classification: The current model performs binary classification (CMB vs. non-CMB). Multi-class classification, which might be more relevant for diagnosing different types or severities of brain lesions, is desired.
  • Accuracy: While high, the 90.05% accuracy is not perfect.
  • Future Work: The authors plan to collect more data to re-test the system, explore its application to detect specific brain diseases, and develop multi-class detection systems. They also intend to investigate advanced deep learning models and structures, such as dilated convolution and AlphaMEX Global Pool.

Summary updated

Abstract

Cerebral microbleed (CMB) is a serious public health concern. It is associated with dementia, which can be detected with brain magnetic resonance image (MRI). CMBs often appear as tiny round dots on MRIs, and they can be spotted anywhere over brain. Therefore, manual inspection is tedious and lengthy, and the results are often short in reproducible. In this paper, a novel automatic CMB diagnosis method was proposed based on deep learning and optimization algorithms, which used the brain MRI as the input and output the diagnosis results as CMB and non-CMB. Firstly, sliding window processing was employed to generate the dataset from brain MRIs. Then, a pre-trained VGG was employed to obtain the image features from the dataset. Finally, an ELM was trained by Gaussian-map bat algorithm (GBA) for identification. Results showed that the proposed method VGG-ELM-GBA provided better generalization performance than several state-of-the-art approaches.
BibTeX
@article{lu2023diagnosis,
  author = {Siyuan Lu and Kaijian Xia and Shui-Hua Wang},
  title = {{Diagnosis of cerebral microbleed via VGG and extreme learning machine trained by Gaussian map bat algorithm}},
  journal = {J. Ambient Intell. Humaniz. Comput.},
  year = {2023},
  url = {https://openreview.net/forum?id=zKK3yzjXHR},
}
Figure