Novel Use of Deep Convolution Architecture Pre-Trained on Surface Crack Dataset to Localize and Segment Wrist Bone Fractures

Abstract

In this day and age, X-rays are the principal instruments for assessing suspected fractures in humans. Expert radiologists are required to suspect the fractures by manually inspecting them, which is a time-consuming process. Automatic detection is beneficial, especially in under-resourced areas where scarce resources and experienced radiologists are observed. Wrist Fracture Dataset (WFD) and Surface Crack Dataset (SCD) were developed to detect and segment wrist bone fractures automatically. The number of wrist fracture images obtained from the Indian hospitals is 315, having 733 annotations/cracks, which is insufficient to produce accurate results using deep learning techniques. As a result, we included SCD for improved model generalization. WFD consists of 3,000 images collected by capturing the minute cracks from road, pavement, and walls, which has similar patterns as the bone fracture cracks. The proposed architecture is a modified version of mask-RCNN architecture where the surface crack dataset's weights are transferred to the wrist X-ray dataset for better model convergence. The results obtained from the modification done at the sub-architecture level (levels 1 and 2) are examined. Combining the modifications proposed at level 1 and level 2, we have obtained improved results against the standard mask-RCNN model for the wrist fracture dataset. We achieved an average precision of 92.278% and 79.003% for fracture detection and 77.445 and 52.156% for fracture segmentation on 50 0 and 75 0 scales, respectively.