Abstract
Dynamic contrast-enhanced fluorescence imaging (DCE-FI) classification of tissue viability in twelve adult patients undergoing below knee leg amputation is presented. During amputation and with the distal bone exposed, indocyanine green contrast-enhanced images were acquired sequentially during baseline, following transverse osteotomy and following periosteal stripping, offering a uniquely well-controlled fluorescence dataset. An unsupervised classification machine leveraging 21 different spatiotemporal features was trained and evaluated by cross-validation in 3.5 million regions-of-interest obtained from 9 patients, demonstrating accurate stratification into normal, suspicious, and compromised regions. The machine learning (ML) approach also outperformed the standard method of using fluorescence intensity only to evaluate tissue perfusion by a two-fold increase in accuracy. The generalizability of the machine was evaluated in image series acquired in an additional three patients, confirming the stability of the model and ability to sort future patient image-sets into viability categories.