Abstract:
Hemophilia is a rare genetic bleeding disorder where blood does not clot properly due
to lack of clotting factors, Factor VIII in Hemophilia A and Factor IX in Hemophilia B,
which can result in spontaneous and recurrent internal bleeding, particularly into
joint spaces such as knees, ankles, and elbows. If not treated quickly, joint bleeding
can lead to pain and swelling, and eventually joint damage. In Bangladesh, most
patients experience prohibitive delay in clinical evaluation, including travel to referral
centers. This paper presents a Novel Real-Time Joint Bleeding Detection and Clinical
Decision Support Tool for the Remote Assessment of Hemophilia-A patient's joint.
Through the Regional Youth Committee, a dataset of more than 2000 images of joint
bleeding has been collected across the country with the approval of Hemophilia Society of
Bangladesh. After Augmentation in Roboflow the image increase to 5000 Certified
hemophilia specialists categorized the images into five classes: severe, moderate, mild,
fixed joint, and no bleeding. Different deep learning models including CNN with Xception
and hybrid model ViTForImageClassification with DenseNet121 were trained. The highest
validation accuracy attained was 81.82% with unedited images, underscoring the
significance of background context in the medical image classification process. The bestperforming model was saved as an .h5 file and this was used to develop a web application
with Flask. Using this system, patients are able to receive on-the-go assessments and
treatment recommendations from doctors in real-time, enabling timely action and
reducing.
