Public Health 2 الصحة العامة

Permanent URI for this collection

https://dspace.alquds.edu/handle/20.500.12213/1060

Browse

Browsing Public Health 2 الصحة العامة by Author "Aesha Loay Ebrahim Enairat"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    A Systematic Comparative Evaluation of Classical Machine Learning Algorithms for Liver Lesion Classification in Ultrasound Imaging
    (Al- Quds University, 2026-01-14) Aesha Loay Ebrahim Enairat; عائشه لؤي ابراهيم انعيرات
    Liver ultrasound is very popular as it is accessible, safe and inexpensive, but the distinction between benign and malignant liver lesions is still difficult to make, owing to the noise, low contrast, and the overlap of the lesions. The thesis is an evaluation of the performance and generalization of classical machine learning techniques in liver lesion classification with ultrasound images in 3 binary tasks, where one of them is benign-normal, another one is malignant-normal, and the last one is benign-malignant. We merged a localized clinical dataset with a publicly available dataset in Zenodo, which created an original set of 6,791 ultrasound images. By eliminating the duplicate and very similar images to avoid redundancy we were left with a curated set of unique images numbered 2,387. We compared the default preprocessing with contrast enhancement with Contrast Limited Adaptive Histogram Equalization (CLAHE) and tested various traditional classifiers, including ensemble model, support vector machines, linear model, instance-based model as well as probabilistic models. Stratified cross-validation, independent testing and a fully isolated holdout set were used in the evaluation of model performance. The best and consistent performance of the ensemble-based classifiers was observed especially when it came to malignant- normal and benign- normal classification. Conversely, benign-malignant classification was the hardest to carry out because the lesion types had large visual overlaps. The use of CLAHE resulted in better sensitivity and lesion separability of a number of models, and task-specific advantages. On the whole, the findings suggest that classical machine learning pipelines with the proper support of preprocessing and strict validation may be effective in terms of assisting the classification of ultrasound-based liver lesions and may serve as the basis of further advancement.