Neural Networks Models for Multi-Attribute Assessment of Fine Grained Wound Images

Liu, Ziyang

Etd

Neural Networks Models for Multi-Attribute Assessment of Fine Grained Wound Images

Public Deposited

Chronic wounds affect 6.5 million Americans and 15 percent of medicare patients. Many wound patients are treated by nurses who visit their homes periodically. Care of chronic wounds involves cleaning, debridement, changing of dressings and applying medicines. However, several problems present a challenge in home settings. Wound measurement, healing progress assessment and documentation can take up to 25 minutes of an hour-long care session. Moreover, human error can also occur. Due to the large and growing number of chronic wounds, there is an increasing demand for more efficient chronic wound care, especially information technology solutions that support the work of medical personnel and reduces the cost of care. Smartphone-based image analyses have emerged as a viable option for remote wound assessment. Our team has previously proposed and developed the Smartphone Wound Analysis and Decision-Support (SmartWAnDS), which autonomously analyzes chronic wound images captured using smartphone cameras and provides wound care recommendations to patients and their caregivers. A visiting nurse simply takes a picture of the patient's wound in their home, and submits it to the SmartWAnDS system for analysis. A key benefit of SmartWanDS is that patients receive standardized assessment and feedback on their wounds. This dissertation focuses on researching developing the SmartWAnDS module that autonomously grades the healing progress of four wound types (diabetic ulcers, pressure ulcers, vascular ulcers and surgical wounds) from their visual appearance in a smartphone photograph. Novel neural networks-based solutions are proposed for three specific research objectives: 1) Multi-attribute, comprehensive wound assessments from smartphone images: A DenseNet Convolutional Neural Network (CNN) framework with patch-based context-preserving attention is proposed to assess all eight PWAT attributes. 2) Wound infection and ischemia detection: a Diabetic Foot Ulcer (DFU) dataset was augmented using geometric and color image operations, after which binary infection and ischemia classification was done using the EfficientNet deep learning model. 3) Robust multi-attribute wound assessment using small, imbalanced, wound dataset: A Semi-Supervised learning and Progressive Multi-Granularity training mechanism were augment a small primary labeled dataset using a secondary corpus of unlabeled wound images. Multi-attribute wound scoring utilized the EfficientNet CNN on the augmented wound corpus. Multi-attribute wound assessment ground truth labels were generated using the clinically-validated Photographic Wound Assessment Tool (PWAT), which assesses eight wound attributes: Size, Depth, Necrotic Tissue Type, Necrotic Tissue Amount, Granulation Tissue type, Granulation Tissue Amount, Edges, Periulcer Skin Viability. All proposed methods were rigorously evaluated and compared to baselines using machine learning classification metrics. The proposed SmartWAnDS system is the first intelligent system that autonomously grades wounds based on the eight criteria in the PWAT rubric, alleviating the significant burden that manual wound grading imposes on wound care nurses.

Creator