College of Science and Technology

Permanent URI for this communityhttp://itsupport.cu.edu.ng:4000/handle/123456789/28738

Welcome to the community of research works in Science and Technology

Browse

Subject: search.filters.subject.Deep Learning

Search Results

Now showing 1 - 5 of 5
  • No Thumbnail Available
    Item
    DEVELOPMENT OF A TRANSFER LEARNING PIPELINE FOR PROSTATE CANCER AGGRESSIVENESS CLASSIFICATION
    (Covenant University Ota, 2025-08) OLUSUYI, Fiyinfoluwa Ruth; Covenant University Dissertation
    Prostate cancer is a leading malignancy in men, where accurate aggressiveness assessment is crucial for guiding treatment. While multi-parametric MRI (mpMRI) is now the established standard for non-invasive diagnosis, its interpretation can be subjective. Deep learning has shown promise, but limited data poses a challenge. This study addresses this limitation by developing a comprehensive transfer learning pipeline for automated prostate cancer aggressiveness classification using mpMRI data. The public PROSTATEx dataset was processed into 2D image patches combining T2-weighted, ADC, and high b-value DWI sequences as 3-channel inputs. Seven state-of-the-art pre-trained Convolutional Neural Network (CNN) architectures, including EfficientNet-B0, ResNet18, VGG16, DenseNet121, MobileNetV3, InceptionV3, and ShuffleNet V2, were fine-tuned using a consistent framework incorporating WeightedRandomSampler and regularization to address class imbalance. Performance evaluation was carried out on a separate validation set using a range of standard metrics, including accuracy, F1-score, specificity, and AUC. The findings identified EfficientNet-B0 as the superior architecture. It delivered the best performance, achieving an overall accuracy of nearly 97% and a macro F1-score of 0.96. This result highlights the exceptional effectiveness of modern, efficient network designs. Remarkably, the lightweight MobileNetV3 delivered nearly identical performance, also achieving a 96% accuracy and macro F1-score. Other architectures, including ShuffleNet V2, DenseNet121, and ResNet18, also proved highly effective with accuracies between 94-96%. The VGG16 and InceptionV3 models did not reach the same level of performance as the leading architectures, with accuracies of 0.72 and 0.90, respectively.
  • No Thumbnail Available
    Item
    A COMPUTATIONAL FRAMEWORK FOR PREDICTING COMPOUNDPROTEIN INTERACTION FOR PROSTATE CANCER THERAPEUTIC DISCOVERY
    (Covenant University Ota, 2025-08) AGBI, Mayowa; Covenant University Dissertation
    Prostate cancer (PCa) is a major public health issue globally. In sub-Saharan Africa, with its limited number of diagnostic and treatment resources, it accounts for high mortality. The conventional approach to drug discovery is lengthy, expensive, and often insufficient to address the complex treatment-resistant prostate cancers present. In this study, a deep learning computational framework to predict Compound-Protein Interactions (CPI) for prostate cancer drug discovery was developed. An end-to-end machine learning pipeline was implemented using curated datasets from Zenodo, ChEMBL, BindingDB, and UniProt. Molecular representations for compounds were constructed using 2048-bit Morgan fingerprints, dimensionally reduced to 200 via Principal Component Analysis (PCA), and for the proteins, 100-dimensional 3-mer Word2Vec embeddings were used. These features were fed into a double-input deep neural network that was optimized with binary-cross-entropy loss, the Adam optimizer, and dropout regularization. The model identified five novel bioactive compounds for targeting proteins of prostate cancer biomarkers. Model confidence was used to prioritize predicted interactions for AR, SRC, and EGFR. Molecular docking in PyRx and AutoDock Vina, followed by visualization in Discovery Studio supporting strong binding affinity (-7.2 to -10) and complementarity from the structural point of view, constituting therapeutic potential. An integration of molecular docking enriched translational value to the prediction. The results presented here point to a disease-specific platform for in silico drug discovery in prostate cancer. This study opens a very promising path toward giving priority to candidate compounds by coupling the deep learning with structure-based affirmation. It provides a very viable ground to be merged with experimental validation and combinatorial therapy design, thereby taking one step further into machine learning-assisted precision oncology.
  • No Thumbnail Available
    Item
    ADAPTING MOBILESAM FOR FEW-SHOT SEGMENTATION OF PROSTATE CANCER IN HISTOPATHOLOGY IMAGES
    (Covenant University Ota, 2025-08) ANTHONY, Micheal IdediA; Covenant University Dissertation
    Segmenting prostate cancer in tissue images is difficult because of irregular gland shapes, broken tissue structures, and very few labelled images available for training. This study introduces FrozenSE-SAM, a segmentation method that works well even with small datasets. It combines a frozen MobileSAM encoder with a lightweight decoder enhanced by Squeeze-and-Excitation (SE) blocks and is trained using Focal Tversky Loss, which helps focus on difficult regions. Unlike older methods that need extra shape information or lots of labels, FrozenSE-SAM can directly segment tumour regions without prompts. It was trained on only 35 tissue microarray (TMA) cores from the Gleason 2019 dataset and tested on 100 new samples. The model achieved a Dice score of 68.45%, which is better than U-Net (60.72%), Swin-UNETR (58.12%), and a Signed Distance Function (SDF) based model (62.77%). For measuring boundary accuracy, FrozenSE-SAM showed better performance with HD95 = 0.0228 mm and ASD = 0.0056 mm, compared to the SDF model (HD95 = 0.0328 mm, ASD = 0.0072 mm), and worse scores from U-Net and Swin-UNETR. Visual/Qualitative result also confirmed that FrozenSE-SAM was better at outlining complex tumour regions. It could accurately segment cribriform and fused glands without including nearby healthy tissue. In contrast, the SDF model produced blurry edges and missed finer structures, leading to under-segmentation. These results show that FrozenSE-SAM is a strong, reliable method for prostate cancer segmentation, especially in real-world situations with limited data.
  • No Thumbnail Available
    Item
    DEVELOPMENT OF A TRANSFER LEARNING PIPELINE FOR PROSTATE CANCER AGGRESSIVENESS CLASSIFICATION
    (Covenant University Ota, 2025-08) OLUSUYI, Fiyinfoluwa Ruth; Covenant University Dissertation
    Prostate cancer is a leading malignancy in men, where accurate aggressiveness assessment is crucial for guiding treatment. While multi-parametric MRI (mpMRI) is now the established standard for non-invasive diagnosis, its interpretation can be subjective. Deep learning has shown promise, but limited data poses a challenge. This study addresses this limitation by developing a comprehensive transfer learning pipeline for automated prostate cancer aggressiveness classification using mpMRI data. The public PROSTATEx dataset was processed into 2D image patches combining T2-weighted, ADC, and high b-value DWI sequences as 3-channel inputs. Seven state-of-the-art pre-trained Convolutional Neural Network (CNN) architectures, including EfficientNet-B0, ResNet18, VGG16, DenseNet121, MobileNetV3, InceptionV3, and ShuffleNet V2, were fine-tuned using a consistent framework incorporating WeightedRandomSampler and regularization to address class imbalance. Performance evaluation was carried out on a separate validation set using a range of standard metrics, including accuracy, F1-score, specificity, and AUC. The findings identified EfficientNet-B0 as the superior architecture. It delivered the best performance, achieving an overall accuracy of nearly 97% and a macro F1-score of 0.96. This result highlights the exceptional effectiveness of modern, efficient network designs. Remarkably, the lightweight MobileNetV3 delivered nearly identical performance, also achieving a 96% accuracy and macro F1-score. Other architectures, including ShuffleNet V2, DenseNet121, and ResNet18, also proved highly effective with accuracies between 94-96%. The VGG16 and InceptionV3 models did not reach the same level of performance as the leading architectures, with accuracies of 0.72 and 0.90, respectively.
  • No Thumbnail Available
    Item
    A MULTI-DOCUMENT SUMMARIZATION APPROACH FOR QUERY-DRIVEN NON-FACTOID QUESTION-ANSWERING SYSTEM
    (Covenant University Ota, 2025-07) EFOSA-ZUWA, Emmanuel Temidire; Covenant University Dissertation
    In Natural Language Processing (NLP), Question Answering Systems (QAS) are essential for facilitating efficient access to relevant information. Traditional QAS approaches typically involve decomposing user queries, retrieving relevant documents, and ranking potential answers, often struggle with non-factoid questions that require detailed, context-rich responses synthesized from multiple sources. While existing research has focused heavily on passage selection and ranking, many methods fail to produce a coherent answer, leaving the challenge of multi-source summarization largely unresolved. This study presents a transfer learning-based QAS framework that addresses non-factoid queries through multi-source summarization. The framework follows a multi-stage methodology incorporating question paraphrasing, contradiction detection, sentence embedding and pruning, and a hybrid approach combining extractive and abstractive summarization techniques. Quantitative and qualitative analyses were conducted using benchmark datasets, including WikiHow QA and PubMedQA to evaluate its effectiveness. The proposed system achieved strong quantitative results, with scores on WikiHow QA (ROUGE-1: 34.10, ROUGE-2: 12.30, ROUGE-L: 32.10, BLEU: 25.14, BERTScore: 95.17) and PubMedQA (ROUGE-1: 42.30, ROUGE-2: 16.10, ROUGE-L: 33.40, BLEU: 31.66, BERTScore: 95.72), demonstrating its ability to generate accurate and contextually relevant answers. Qualitative evaluations also yielded promising outcomes, with average ratings of 4.37 for information, 4.16 for conciseness, 4.20 for readability, and 4.01 for correctness on a 5-point scale, confirming the model’s effectiveness in delivering accurate and comprehensible responses. This transfer learning-based QAS framework contributes meaningfully to advancements in NLP and offers valuable support for researchers and developers working on intelligent, explainable, and practical question answering systems.