In 2021, 1 out of adults suffered from diabetes mellitus globally. A major complication resulting from this is diabetic foot ulcers (DFU) , which can lead to amputation if improperly managed. Current clinical approaches to DFU have significant limitations, such as the high cost involved in the diagnosis and time-consuming care. In light of this, we have developed DFUCare, a mobile app that uses the built-in camera and deep learning to diagnose and quantify patients' diabetic foot ulcers. DFUCare is trained on an extensive foot dataset containing over 5000 images of normal and infected DFU patients. For binary classification of infected and non-infected patients, we use a combination of transfer and ensemble learning. Using transfer learning, DFUCare uses the pretrained InceptionResNetV2 model for deep learning derived feature extraction. With an ensemble approach, we compare the deep learning derived feature extraction results to handcrafted color and textural feature extraction results to come up with a final prediction of infected vs. non-infected patients. In the end, we achieved top-level results with a binary accuracy of .9014, AUC of .9437, specificity of .9023, and precision of .8901. In addition, using standardized metrics and computer vision algorithms, DFUCare can determine the dimensions of the DFU to quantify the severity over time. With all of the information stored in the cloud, doctors and patients can use the share feature in the mobile app to send medical reports to anyone in the world. In the future, we are planning to add clinical, biological, and epidemiological features along with macroscopic image features to improve the performance of the tool in predicting infection and healing. This novel approach has the potential to deliver a paradigm shift in diabetic foot care among patients by being a cost-effective, remote, and convenient healthcare solution. Disclosure V.Sendilraj: None. W.C.Pilcher: None. M.K.Bhasin: None.
Read full abstract