M.Phil / MS
Permanent URI for this collectionhttps://repository.cuilahore.edu.pk/handle/123456789/36
This collection archives the complete set of theses produced by students of the COMSATS University Islamabad, Lahore Campus.
Browse
5 results
Search Results
Item Deep Learning Based Improved Endometrial Cancer Detection Using Comprehensive MRI Dataset(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Saba Khan; CIIT/SP23-RCS026/LHR; Dr. Zulfiqar Habib; LHR TP 10032Endometrial cancer (EC) needs the correct diagnosis of this cancer during the initial stage, with uterine and tumor boundaries, and magnetic resonance imaging (MRI) represents the clinical gold standard. Nevertheless, all MRI-based EC investigations have three significant limitations that include the unavailability of publicly accessible datasets, the usage of single-sequence or single-plane magnetic resonance imaging, and the unavailability of strong deep learning models to segment the uterus and tumor consistently. In order to fill in these gaps, the proposed paper develops a deep learning model that is a hybrid ResNet-34 + Attention U-Net model, which employs the full diagnostic capabilities of a combination of multi-sequence (nine MRI sequences) and multi-plane (axial, sagittal, and coronal) MRI. One diagnostic center (20212024) was used to construct a large and well-curated local dataset (LD), including normal and stage I EC cases, and two external ones provided by Mao et al. (2022) (MD) and Xiong et al. (2023) (XD) were used only to cross-verify ecological datasets. The decoder has attention gates that boost anatomically weighty areas without reducing the background noises, permitting the precise 3-class division of the background, uterus, and tumor. There were systematic experiments of several encoder-decoder designs, attention, and loss functions trained according to a single training protocol with hyperparameters optimized by Optuna. The proposed model produced the state-of-the-art performance with the overall accuracy of 0.988 on the local dataset, surpassing Mao et al. (0.914) and Xiong et al. (0.869), and a higher Dice coefficient, IoU, precision, recall, and F1-score. Cross-dataset testing also indicated good generalization, with the model proposed showing consistently high accuracy on the MD and XD test sets, which proved that it is strong at different institutions and imaging protocols. The findings validate the proposed structure as a clinically reliable, reproducible, and generalizable solution to EC segmentation at an early stage.Item 3D Surface Reconstruction using Point Cloud Segmentation with Machine Learning(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2020) Naveed Ahsan Ali Raza; FA17-RCS-018; LHR TP 5977; Dr. Zulfiqar Habib3D point cloud is a kind of geometric data-structure, which is easily acquired using highly advanced 3D sensors. These sensing devices are capable of producing a considerable amount of 3D point clouds by capturing the 3D surface geometries and dimensions from targeted objects and real-world scenes. The captured 3D point clouds are utilized to reconstruct 3D objects, surfaces, shapes, and parts. 3D surface reconstruction is widely used in reverse engineering, 3D medical imaging, 3D printing, and automated medical surgery. Discrete surface reconstruction from 3D point clouds is still a challenging task, because captured data is un-ordered, noisy, redundant, and has topological imperfections. Due to these issues, it is not adequate for 3D surface reconstruction in its current form. In this direction, techniques based on geometric modeling have already been proposed to solve 3D surface reconstruction problems from point cloud data. Due to the irregular format of 3D point clouds, many researchers convert the data into 3D voxel grids or collection of images, which add unnecessary volume and cause issues. On the other hand, most of the deep learning-based techniques have focused on regular input data for surface reconstruction, and little attention has been paid towards 3D point cloud data. Therefore, to address the issues mentioned above, the proposed model is based on unified deep net architecture. The proposed deep net architecture takes raw 3D point clouds as an input. It performs a segmentation technique to tackle un-ordered data and further utilizes this segmentation information for 3D surface reconstruction of objects and their parts. To evaluate the performance of proposed model, experiments have been performed on the benchmark ShapeNet dataset. This dataset is a large-scale repository of 3D point clouds containing 16,880 objects of 16 different categories. The results show that the proposed method achieves superior results as compared to state-of-the art methods with an accuracy of 82% for 3D surface reconstructionItem Anchor Free Motorbike Detection in Surveillance Videos of Dense Traffic(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2020) Mahwish Irshad; FA16-RCS-029; LHR TP 5972; Dr. Zulfiqar HabibObject detection and tracking have become the most significant and challenging task within the field of Computer Vision that attempts to detect, recognize, and track objects over any sequence of frames (images) called video. Its purpose is to locate objects motion in a video file or surveillance camera. Object detection is the procedure of locating one or multiple objects by utilizing a single camera, multiple cameras, or a provided video file. Urban area expansions have increased the demand for proper traffic surveillance. Traffic monitoring is important for detecting road accidents, collecting evidence for the investigation, tracking criminals, and traffic violators. To meet these requirements, intelligent systems can be developed to extract and analyze traffic information. This research work is concerned with the detection of motorbikes in videos. Detection of motorbikes object is very helpful for organizations and investigators, which are mostly concerned in dealing with potential violations of human rights on roads. For the detection of motorbike objects a state-of-the-art dataset called Velastin7500 is used. The thesis focused on the detection of motorbikes from videos by using two pre-trained, RetinaNet, and YOLOv3 architectures on the Velastin7500 dataset to explore the importance of object detection from images, captured by a drone camera. This research work has also inspected the domain shift issue of learning of features from images to detect objects and by pertaining this information to real-world imagery in the context of detection of motorbike. The proposed work also explored serious issues in the existing algorithm based on anchor free technique for some practical applications. The results produced for the Velastin7500 dataset by implementing above mentioned two architectures of RetinaNet and YOLOv3 have mean average precision (mAP) value of 18.25% and 76.5%, respectively. This research will also be helpful to understand why the anchor free approach did not work well with partially visible objects.Item Detecting Breast Cancer from Histopathology Images Using CNN-Based Framework(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Tahreem Zara; CIIT/SP23-RCS-028/LHR; Dr. Zulfiqar Habib; LHR TP 9706Breast cancer remains one of the most frequently diagnosed cancers among women globally, with histopathology analysis serving as the gold standard for diagnosis. But manual examination of histopathology slides takes a long time, is prone to making mistakes, and often has differences between observers, especially in complicated cases. To overcome these constraints, this study introduces a lightweight and deployable deep learning framework designed to automate breast cancer detection utilizing histopathology images from the BRACS dataset. The proposed framework utilizes the DenseNet169 model pre-trained on ImageNet; we fine-tuned the last 30 layers after testing different depths. We resized the input images to 1024×1024 to preserve important morphological details, and we employed simple augmentations, such as rotation and flipping, during training. Additionally, a semi-automated segmentation and annotation strategy was developed using k-means clustering based on RGB color and shape features, generating binary masks that were further validated by a histopathologist. Both classification and segmentation models were implemented in Keras on a CPU-based environment with 32 GB RAM. We evaluated our models using scikit-learn. The classification achieved 79% accuracy on the 3-class test set and 90% accuracy on the binary test set (normal vs. invasive carcinoma), demonstrating competitive results compared to state-of-the-art models. The evaluation metrics used for both models were precision, recall, and F1-score; classification performance was further assessed using accuracy, and segmentation performance was evaluated using the Dice coefficient. Our classification model (DenseNet169 with partial fine-tuning) has approximately 4 million trainable parameters, while the segmentation model (U-Net with ResNet34 encoder) consists of around 25 million parameters—both models being compact enough for deployment on standard desktop CPUs. The segmentation masks were generated using a novel semi-automated pipeline that balances visual coherence (shape and structure) with RGB clustering, allowing more precise annotation than conventional thresholding or fully automated tools. These findings highlight that a selectively fine-tuned CNN can deliver reliable diagnostic assistance, even in resource- constrained environments.Item Development of an Advanced Machine Learning Model for 3D Brain Segmentation in MRI(Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Anam Amjad; CIIT/SP23-RCS-007/LHR; Dr. Zulfiqar Habib; LHR TP 9701In the last few years, the rates of occurrence and death associated with brain tumors have shown a worrisome increase and now appear to be on a sustained uptick. The most recent report from the American Cancer Society (ACS) estimated that in 2020, 24,531 people in the United States would receive a diagnosis of a brain or other central nervous system tumor, and 18,306 would die from one of those tumors. The current research effort is a penetration into the brain tumor segmentation problem, and it seems to kick off with a quite elaborate preamble. To solve the problems of the heterogeneous tumors, the inconsistent morphology, and the deployment of clinical models, we propose a revised model of 3D Residual U-Net. This model has incorporated Convolutional Block Attention Modules (CBAM) a type of neural network component that allows it to better understand which pixels to pay attention to when outputting its results. We trained and evaluated this model on the BraTS benchmark 2020 dataset. The suggested framework attains high segmentation accuracy, along with computational efficiency, amassing Dice scores of 68% for ET, 83% for TC, and 91% for WT, strong generalization to the diverse, complex morphologies of brain tumors being evident here. By amending U-Net with residual learning, attention mechanisms, and a composite loss function (Dice + Focal), this framework provides a clinically viable approach that bridges the gap between research and real-world applicability in neuro- oncological imaging.