A study on assessing classifier accuracy in detecting intracranial hemorrhage using transfer learning models on ct scan images /
Amiir Haamzah Mohamed Ismail,
A study on assessing classifier accuracy in detecting intracranial hemorrhage using transfer learning models on ct scan images / Amiir Haamzah Mohamed Ismail - xiv, 122 pages : illustrations (some color) ; 30 cm. + 1 CD-COM
Faculty of Manufacturing and Mechatronic Engineering Technology
Thesis (Master of Science) -- Universiti Malaysia Pahang – 2026
Includes bibliographical references
Intracranial hemorrhage (ICH) is a life-threatening condition in which the Malaysia population has been reported with one year mortality rate of 48.5%. The survival rate of patients depends heavily on rapid and accurate diagnosis. Thus, non-contrast computed tomography (CT) scans remain as the gold standard for detection. However, expert radiological interpretation is crucial, which is often delayed in emergency departments with resource limited settings. AI in particular transfer learning has emerged as a promising approach for ICH detection, but existing studies commonly relied on heavily customized architectures, extensive fine tuning, or proprietary pipelines. These approaches limits reproducibility and make it difficult to quantify the baseline performance that can be achieved with pretrained models used out the box. Consequently, the research problem addressed in this study is the absence of a systematic, standardized, and reproducible comparison of pretrained CNN for binary classification of detecting ICH based on filtered RSNA dataset. This study pursued three objectives, starting with preparing a balanced, stratified, CT scan dataset from publicly available RSNA source. Subsequently, develop a standardized transfer-learning binary classification pipeline that employs 17 pretrained CNN architectures from Keras as fixed feature extractors, couples them with four classical classifiers, and optimizes classifier hyperparameters through randomized search. Finally, benchmark the resulting model-classifier combinations using several performance metrics. From a methodology standpoint, an even sample size of 2,000 images, with equal numbers of ICH and non-ICH cases, was randomly selected from the RSNA dataset, and split into a stratified 70/15/15 training/validation/testing split. Pre-processing included standardizing all images in a brain windowing mode (level 40 Hounsfield units, width 80 Hounsfield units), and resizing images to their corresponding native resolutions. All 17 model architectures tested belonged to one of the following families: VGG (VGG16, VGG19), ResNet (ResNet50, 101, 152, and their V1/V2 variants), Inception (InceptionV3, InceptionResNetV2, Xception), Mobile (MobileNet, MobileNetV2), DenseNet (DenseNet121, 169, 201), and NAS (NASNetMobile). The final classification performed through each of four classical machine learning algorithms: Logistic Regression, Support Vector Machine, Random Forest, and k-Nearest Neighbour. The results indicate that ICH identification is indeed possible using pre-trained models as fixed feature extractors without the hassle of fine-tuning, and the importance of architecture design over number of parameters in facilitating transfer learning. With a peak accuracy of 76.33%, the results provide an empirical sound benchmark to gauge the added value of fine-tuning, multi-window image processing, and ensemble learning. Moreover, the results suggest that efficient models are capable of being applied clinically, especially in less developed facilities, such as regional hospitals and mobile devices. Although the current performance level is not high enough for clinical deployment, it is suitable for acting as an auxiliary tool that marks suspicious cases requiring radiologist examination. Future research should focus on fine-tuning techniques, multi-class classification of ICH subtypes, cross-validation of models across multiple datasets from different institutions, and clinical applications of models.
THE0010812(Local)
Faculty of Manufacturing and Mechatronic Engineering Technology--Dissertations
Universities and colleges--Dissertations
Thesis--Dissertations
A study on assessing classifier accuracy in detecting intracranial hemorrhage using transfer learning models on ct scan images / Amiir Haamzah Mohamed Ismail - xiv, 122 pages : illustrations (some color) ; 30 cm. + 1 CD-COM
Faculty of Manufacturing and Mechatronic Engineering Technology
Thesis (Master of Science) -- Universiti Malaysia Pahang – 2026
Includes bibliographical references
Intracranial hemorrhage (ICH) is a life-threatening condition in which the Malaysia population has been reported with one year mortality rate of 48.5%. The survival rate of patients depends heavily on rapid and accurate diagnosis. Thus, non-contrast computed tomography (CT) scans remain as the gold standard for detection. However, expert radiological interpretation is crucial, which is often delayed in emergency departments with resource limited settings. AI in particular transfer learning has emerged as a promising approach for ICH detection, but existing studies commonly relied on heavily customized architectures, extensive fine tuning, or proprietary pipelines. These approaches limits reproducibility and make it difficult to quantify the baseline performance that can be achieved with pretrained models used out the box. Consequently, the research problem addressed in this study is the absence of a systematic, standardized, and reproducible comparison of pretrained CNN for binary classification of detecting ICH based on filtered RSNA dataset. This study pursued three objectives, starting with preparing a balanced, stratified, CT scan dataset from publicly available RSNA source. Subsequently, develop a standardized transfer-learning binary classification pipeline that employs 17 pretrained CNN architectures from Keras as fixed feature extractors, couples them with four classical classifiers, and optimizes classifier hyperparameters through randomized search. Finally, benchmark the resulting model-classifier combinations using several performance metrics. From a methodology standpoint, an even sample size of 2,000 images, with equal numbers of ICH and non-ICH cases, was randomly selected from the RSNA dataset, and split into a stratified 70/15/15 training/validation/testing split. Pre-processing included standardizing all images in a brain windowing mode (level 40 Hounsfield units, width 80 Hounsfield units), and resizing images to their corresponding native resolutions. All 17 model architectures tested belonged to one of the following families: VGG (VGG16, VGG19), ResNet (ResNet50, 101, 152, and their V1/V2 variants), Inception (InceptionV3, InceptionResNetV2, Xception), Mobile (MobileNet, MobileNetV2), DenseNet (DenseNet121, 169, 201), and NAS (NASNetMobile). The final classification performed through each of four classical machine learning algorithms: Logistic Regression, Support Vector Machine, Random Forest, and k-Nearest Neighbour. The results indicate that ICH identification is indeed possible using pre-trained models as fixed feature extractors without the hassle of fine-tuning, and the importance of architecture design over number of parameters in facilitating transfer learning. With a peak accuracy of 76.33%, the results provide an empirical sound benchmark to gauge the added value of fine-tuning, multi-window image processing, and ensemble learning. Moreover, the results suggest that efficient models are capable of being applied clinically, especially in less developed facilities, such as regional hospitals and mobile devices. Although the current performance level is not high enough for clinical deployment, it is suitable for acting as an auxiliary tool that marks suspicious cases requiring radiologist examination. Future research should focus on fine-tuning techniques, multi-class classification of ICH subtypes, cross-validation of models across multiple datasets from different institutions, and clinical applications of models.
THE0010812(Local)
Faculty of Manufacturing and Mechatronic Engineering Technology--Dissertations
Universities and colleges--Dissertations
Thesis--Dissertations