Attention-Based Multiple Instance Learning for Achilles Tendinopathy Detection in Ultrasound Imaging

Achilles tendinopathy is a common condition resulting in pain and inflammation in the Achilles tendon due to injury or overuse. Ultrasound is typically the medical imaging modality of choice during diagnosis due to its relatively broader accessibility, lower cost, and non-invasiveness when compared to other technologies but can produce noisy, poorer quality images depending on many factors in data collection. Past studies have applied fully supervised machine learning methods for detection of Achilles tendinopathy, requiring strenuous data cleaning and labeling, which is time consuming for physicians. To more closely align with a typical clinical workflow, this work applies a weakly supervised multiple instance learning approach with attention-based pooling to detect and emphasize ultrasound images in a set that are more indicative of a patient's tendon health. This approach achieved 97.89% area under the ROC curve and demonstrated a promising capability to ignore images with artifacts and images of lesser quality.