Keras Implementation in Detecting Intracranial Hemorrhage and Multiclass Classification of Subtypes via Transfer Learning and Classifiers Selection
The development of deep neural networks for medical imaging applications, especially the diagnosis of intracranial hemorrhage (ICH) from CT scans, is greatly aided by machine learning frameworks such as Keras. This work investigates a pipeline that uses Keras' neural modules to distinguish betw...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Penerbit UMP
2024
|
| Subjects: | |
| Online Access: | http://umpir.ump.edu.my/id/eprint/43004/1/Keras%20Implementation%20in%20Detecting%20Intracranial%20Hemorrhage.pdf http://umpir.ump.edu.my/id/eprint/43004/ https://doi.org/10.15282/mekatronika.v6i2.11358 https://doi.org/10.15282/mekatronika.v6i2.11358 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The development of deep neural networks for medical imaging applications, especially the diagnosis of intracranial hemorrhage (ICH) from CT scans, is greatly aided by machine learning frameworks such as Keras. This work investigates a pipeline that uses Keras' neural modules to distinguish between CT scans of the normal head and those with ICH. Transfer learning models are then used to categorize ICH subtypes. An extensive analysis of current research and techniques demonstrates the effectiveness of deep learning in medical imaging and emphasizes how AI may improve radiologists' diagnostic precision. Using windowing techniques to improve diagnostic features, the study preprocesses pictures from the RSNA Intracranial Hemorrhage Detection dataset. The study assesses performance indicators such classification accuracy using SVM, k-NN, and Random Forest classifiers combined with built-in models from Keras, such as Xception and DenseNet. Findings show that the Xception-SVM pipeline performs exceptionally well in binary classification tasks, achieving 76.33% accuracy, while DenseNet201-SVM performs well in multiclass classification, achieving 60% accuracy. These results highlight how crucial it is to choose the right pipelines for certain classification jobs in order to achieve the best results possible when using medical image analysis. In order to improve diagnostic precision in identifying cerebral hemorrhages, future research directions include increasing classifier performance, investigating sophisticated preprocessing techniques, and fine-tuning models. |
|---|