ADVERTISEMENT
The Impact of Reducing Hospitalizations and Amputations in Patients With Diabetic Foot Ulcers
Purpose: In the United States, it is estimated that 6.5 million people over the age of 40 have peripheral arterial disease (PAD). Radiologists examine computed tomography angiography (CTA) images for detection of clinically significant stenoses. Maximum intensity projections (MIPs) and angled 2D views of CTAs that have often undergone bone subtraction allow radiologists to study vasculature unimpeded by occlusions. Often, subtle signs of PAD can be underreported. Deep learning methods underpinned on regional attention can augment a radiologist’s ability to identify PAD from standard-of-care CTA runoff projections after applying automated bone subtraction tools such as Autobone & VesselIQ Xpress (GE Healthcare).
Materials and Methods: Patient-specific bone-subtracted MIPs consisted of 7 normal and 11 PAD. The train/validation set consisted of 126 MIPs (bone-subtracted CTA runoff projections at ~ 6-degree rotations) from 4 patients (2 normal, 2 PAD) corresponding to 66 normal and 60 indicating PAD. Each MIP was resampled to 900 x 300 pixels before being split into 3* 300 x 300 segments along the subject’s height (3 segments per MIP image). We implement a vision transformer that: 1) transforms each image segment into 100 square patches, 2) extracts neural representations (fixed length numerical vectors), and 3) culminates in a multilayer perceptron that classifies PAD status.
Results: Segment-level classification was performed on the validation set where each prediction is associated with a segment. Segment-level classification achieved A) train set: 88% accuracy, 93% sensitivity, and 83% specificity; B) validation set: 91% accuracy, 91% sensitivity, and 92% specificity. MIP-level classification was performed on the validation set where each prediction is associated with the majority prediction taken across each MIP’s segments (3 segment predictions and 1 MIP prediction). MIP-level classification performance achieved A) train set: 100% accuracy, 100% sensitivity, and 100% specificity; B) validation set: 92% accuracy, 89% sensitivity, and 95% specificity.
Conclusions: Novel binary classifier has promise for identifying PAD status from MIPs and improving the screening of CTA runoffs during radiologist review. Our 3-segment approach to characterizing MIPs may also have potential for regional localization of PAD.