RETURN TO ABSTRACT LISTING


Background
Hallux valgus deformity results in the deviation of the big toe towards the other toes, leading to the formation of a bony prominence known as a bunion. The initial assessment typically involves a clinical examination, where the foot is examined to assess the degree of deformity. Radiographs then provide a measurement of the hallux valgus angle. This angle is measured by manually drawing lines between specific points on the bones of the big toe and the first metatarsal which determines the severity of the deformity. Since these measurements are typically measured manually, the task can be subjective and time consuming. The integration of artificial intelligence (AI) offers a simple solution for automating and standardizing this process.

Educational Goals / Teaching Points
Describe an approach for developing an AI model that localizes and assesses hallux valgus deformity using standard foot radiographs.

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
Primary training: Data set of five foot x-rays for training phase; subsequent evaluative phase; model is underpinned by a Mask R-CNN framework, incorporating an encoder backbone extrapolated from the ResNet-FPN architecture; and preliminary bounding boxes are then subjected to further refinement via the Mask R-CNN predictor, culminating in the delineation of precise Regions of Interest (ROIs). Data augmentation: Creation of the PyTorch dataset; techniques such as Gaussian noise, Gaussian blur, brightness adjustments, smoothing, rotation, translation, and scaling were employed; images were resized to 768x512 for processing and reverted to their original dimensions postprediction; and the model classifies five regions: background, M1, M2, pp1, and pp2 of the foot. Training structured with the original image and separate images for each bone class with its specific mask. Adam optimizer was chosen over SGD due to its efficiency with smaller datasets. Post localization: Outermost edges of the masks for each bone were identified; pixel locations of these contours were used to determine the min and max positions, generating a long axis; axis was divided into the first 25% and last 75% to find centroids for specific sections, creating a new axis representing the bone of interest; and this process was repeated for each bone, and the angle between bones of interest, specifically M1 and PP1, was calculated. Preliminary results show that the masks generated by the AI model have a DICE score > 95% when compared to the manually segmented image.

Conclusion
Our approach to hallux valgus deformity assessment using AI offers a simple method to localize and evaluate the deformity's severity. The model's ability to accurately identify and measure specific foot bones paves the way for more standardized and efficient radiographic evaluations.