Efficient and Accurate Machine Learning Algorithms Based on Transform Domain Techniques for Biomedical Data
Saved in:
| Title: | Efficient and Accurate Machine Learning Algorithms Based on Transform Domain Techniques for Biomedical Data |
|---|---|
| Authors: | Fasihi, Maedeh Sadat |
| Committee Members: | Mikhael, Wasfy |
| Summary: | Automatic medical image diagnosis has gained a great deal of attention in recent years. Early detection of diseases such as cancer, not only increases the survival rate of patients, but also decreases the medical cost and financial burden on both patients and healthcare system. Most current medical image processing techniques turn towards deep learning methods. These methods often require huge amount of data to be reliable, which is not the case in most medical image applications. In this dissertation, we propose a method for diagnosis of brain tumors based on limited training data. The proposed method extracts some valuable features from the images using Wavelet and Discrete Cosine Transform (DCT) transforms. The selected features are then used to classify the tumors. The main characteristics of the selected features are sparsity and generalizability; which allows the method to be fast and accurate. While utilizing wavelets and DCT domains improves the quality of detection considerably, the method is not limited to these domains only. Another aspect of working with medical image data is that, we normally have a sequence of scans for patients, while having only one label for the whole sequence. Our previous classification method was extended to account for this type of limited labeled data. This method can be categorized as weakly supervised technique in a sense that we have one label for the whole sequence. The method was utilized to classify brain tumor grades and achieved state of the art performance with the resolution of 256 x 256. We also conducted a comprehensive set of experiments to analyze the effect of each component on the performance. In order to maintain the connection between the slices and keep the important information in each slice, we used recurrent neural network network to analyze the features. This network assures that the dependency of extracted features from the slices is kept. We also provided a general but automated method to accurately detect and characterize tissue abnormalities from a single slice of medical scan. This technique not limited to the type of anomalies presented here and can be applied to other types of malignancies. |
| URL: | https://stars.library.ucf.edu/etd2020/1885 |
| Database: | OpenDissertations |
| Abstract: | Automatic medical image diagnosis has gained a great deal of attention in recent years. Early detection of diseases such as cancer, not only increases the survival rate of patients, but also decreases the medical cost and financial burden on both patients and healthcare system. Most current medical image processing techniques turn towards deep learning methods. These methods often require huge amount of data to be reliable, which is not the case in most medical image applications. In this dissertation, we propose a method for diagnosis of brain tumors based on limited training data. The proposed method extracts some valuable features from the images using Wavelet and Discrete Cosine Transform (DCT) transforms. The selected features are then used to classify the tumors. The main characteristics of the selected features are sparsity and generalizability; which allows the method to be fast and accurate. While utilizing wavelets and DCT domains improves the quality of detection considerably, the method is not limited to these domains only. Another aspect of working with medical image data is that, we normally have a sequence of scans for patients, while having only one label for the whole sequence. Our previous classification method was extended to account for this type of limited labeled data. This method can be categorized as weakly supervised technique in a sense that we have one label for the whole sequence. The method was utilized to classify brain tumor grades and achieved state of the art performance with the resolution of 256 x 256. We also conducted a comprehensive set of experiments to analyze the effect of each component on the performance. In order to maintain the connection between the slices and keep the important information in each slice, we used recurrent neural network network to analyze the features. This network assures that the dependency of extracted features from the slices is kept. We also provided a general but automated method to accurately detect and characterize tissue abnormalities from a single slice of medical scan. This technique not limited to the type of anomalies presented here and can be applied to other types of malignancies. |
|---|
