Browsing by Author "Wasswa, William"

Now showing 1 - 2 of 2
  • Thumbnail Image
    ItemOpen Access
    A Multimodal Convolutional Neural Network Based Approach for DICOM Files Classification. , 2025(1), e70107.
    (wiley, 2025) Mabirizi, Vicent; Wasswa, William; Kawuma, Simon
    In this study, we developed a convolutional neural network approach for directly classifying digital imaging and communication in medicine files in medical imaging applications. Existing models require converting this format into other formats like portable network graphics. This conversion leads to metadata loss and classification bias, the developed model processes raw digital imaging and communication in medicine files, thereby preserving both pixel data and embedded metadata. The model was evaluated on chest X-ray images for tuberculosis detection and magnetic resonance imaging scan images for brain tumour classification from the National Institute of Allergy and Infectious Diseases. The X-ray modality achieved a precision of 92.9%, recall of 88.4%, F1-score of 90.6% and accuracy of 90.9%, while the magnetic resonance imaging modality obtained a precision of 80.0%, recall of 79.4%, F1-score of 79.7% and accuracy of 85.5%. These results demonstrate the model’s effectiveness across multiple imaging modalities. A key advantage of this approach is the preservation of diagnostic metadata, enhancing accuracy and reducing classification bias. The study highlights its potential to improve medical imaging and support real-time clinical decision making. Despite the promising results, the study acknowledges limitations in dataset diversity and computational efficiency, with future work focusing on addressing these challenges and further optimising the model for deployment in resource-limited environments.
  • Thumbnail Image
    ItemOpen Access
    Deep Learning Techniques in DICOM Files Classification: A Systematic Review
    (BON VIEW PUBLISHING PTE.LTD., 2025) Mabirizi, Vicent; Kawuma, Simon; Natumanya, Deborah; Wasswa, William
    The digital imaging and communications in medicine (DICOM) format is a widely adopted standard for storing medical imaging data, integrating both image and metadata critical for clinical diagnostics. However, its complexity poses challenges for deep learning applications, particularly in extracting and processing this dual-layered data. This review analyzes 23 peer-reviewed studies published between 2014 and 2024, sourced from PubMed, Google Scholar, PLOS, Science Direct, and IEEE databases. Guided by Arksey and O’Malley’s scoping methodology, the review reveals that existing deep learning techniques typically rely on converting DICOM images into simpler formats like JPEG, TIF, or PNG for classification, a process that often results in metadata loss and reduced classification accuracy. Frameworks such as MONAI, NVIDIA Clare, SimpleITK, and OpenCV facilitate direct DICOM processing but face limitations, including overfitting, challenges with data heterogeneity, and inefficiencies in handling large datasets. This review emphasizes the urgent need for developing a robust convolutional neural network architecture capable of directly processing DICOM data to preserve metadata integrity and enhance predictive performance, paving way for more reliable and scalable medical imaging solutions.