Wensheng Fan

Outlier detection and removal improves accuracy of machine learning approach to multispectral burn diagnostic imaging.

Abstract. Multispectral imaging (MSI) was implemented to develop a burn tissue classification device to assist burn surgeons in planning and performing debridement surgery. To build a classification model via machine learning, training data accurately representing the burn tissue was needed, but assigning raw MSI data to appropriate tissue classes is prone to error. We hypothesized that removing outliers from the training dataset would improve classification accuracy. A swine burn model was developed to build an MSI training database and study an algorithm’s burn tissue classification abilities. After the ground-truth database was generated, we developed a multistage method based on Z-test and univariate analysis to detect and remove outliers from the training dataset. Using 10-fold cross validation, we compared the algorithm’s accuracy when trained with and without the presence of outliers. The outlier detection and removal method reduced the variance of the training data. Test accuracy was improved from 63% to 76%, matching the accuracy of clinical judgment of expert burn surgeons, the current gold standard in burn injury assessment. Given that there are few surgeons and facilities specializing in burn care, this technology may improve the standard of burn care for patients without access to specialized facilities.

Surgical wound debridement sequentially characterized in a porcine burn model with multispectral imaging.

INTRODUCTION Multispectral imaging (MSI) is an optical technique that measures specific wavelengths of light reflected from wound site tissue to determine the severity of burn wounds. A rapid MSI device to measure burn depth and guide debridement will improve clinical decision making and diagnoses. METHODOLOGY We used a porcine burn model to study partial thickness burns of varying severity. We made eight 4 × 4 cm burns on the dorsum of one minipig. Four burns were studied intact, and four burns underwent serial tangential excision. We imaged the burn sites with 400-1000 nm wavelengths. RESULTS Histology confirmed that we achieved various partial thickness burns. Analysis of spectral images show that MSI detects significant variations in the spectral profiles of healthy tissue, superficial partial thickness burns, and deep partial thickness burns. The absorbance spectra of 515, 542, 629, and 669 nm were the most accurate in distinguishing superficial from deep partial thickness burns, while the absorbance spectra of 972 nm was the most accurate in guiding the debridement process. CONCLUSION The ability to distinguish between partial thickness burns of varying severity to assess whether a patient requires surgery could be improved with an MSI device in a clinical setting.

The importance of illumination in a non-contact photoplethysmography imaging system for burn wound assessment

We present a non-contact, reflective photoplethysmogram (PPG) imaging method and a prototype system for identifying the presence of dermal burn wounds during a burn debridement surgery. This system aims to provide assistance to clinicians and surgeons in the process of dermal wound management and wound triage decisions. We examined the system variables of illumination uniformity and intensity and present our findings. An LED array, a tungsten light source, and eventually high-power LED emitters were studied as illumination methods for our PPG imaging device. These three different illumination sources were tested in a controlled tissue phantom model and an animal burn model. We found that the low heat and even illumination pattern using high power LED emitters provided a substantial improvement to the collected PPG signal in our animal burn model. These improvements allow the PPG signal from different pixels to be comparable in both time-domain and frequency-domain, simplify the illumination subsystem complexity, and remove the necessity of using high dynamic range cameras. Through the burn model output comparison, such as the blood volume in animal burn data and controlled tissue phantom model, our optical improvements have led to more clinically applicable images to aid in burn assessment.

Multispectral and Photoplethysmography Optical Imaging Techniques Identify Important Tissue Characteristics in an Animal Model of Tangential Burn Excision.

Burn excision, a difficult technique owing to the training required to identify the extent and depth of injury, will benefit from a tool that can cue the surgeon as to where and how much to resect. We explored two rapid and noninvasive optical imaging techniques in their ability to identify burn tissue from the viable wound bed using an animal model of tangential burn excision. Photoplethysmography (PPG) imaging and multispectral imaging (MSI) were used to image the initial, intermediate, and final stages of burn excision of a deep partial-thickness burn. PPG imaging maps blood flow in the skin’s microcirculation, and MSI collects the tissue reflectance spectrum in visible and infrared wavelengths of light to classify tissue based on a reference library. A porcine deep partial-thickness burn model was generated and serial tangential excision accomplished with an electric dermatome set to 1.0 mm depth. Excised eschar was stained with hematoxylin and eosin to determine the extent of burn remaining at each excision depth. We confirmed that the PPG imaging device showed significantly less blood flow where burn tissue was present, and the MSI method could delineate burn tissue in the wound bed from the viable wound bed. These results were confirmed independently by a histological analysis. We found these devices can identify the proper depth of excision, and their images could cue a surgeon as to the preparedness of the wound bed for grafting. These image outputs are expected to facilitate clinical judgment in the operating room.

Burn injury diagnostic imaging device's accuracy improved by outlier detection and removal

Multispectral imaging (MSI) was implemented to develop a burn diagnostic device that will assist burn surgeons in planning and performing burn debridement surgery by classifying burn tissue. In order to build a burn classification model, training data that accurately represents the burn tissue is needed. Acquiring accurate training data is difficult, in part because the labeling of raw MSI data to the appropriate tissue classes is prone to errors. We hypothesized that these difficulties could be surmounted by removing outliers from the training dataset, leading to an improvement in the classification accuracy. A swine burn model was developed to build an initial MSI training database and study an algorithm’s ability to classify clinically important tissues present in a burn injury. Once the ground-truth database was generated from the swine images, we then developed a multi-stage method based on Z-test and univariate analysis to detect and remove outliers from the training dataset. Using 10-fold cross validation, we compared the algorithm’s accuracy when trained with and without the presence of outliers. The outlier detection and removal method reduced the variance of the training data from wavelength space, and test accuracy was improved from 63% to 76%. Establishing this simple method of conditioning for the training data improved the accuracy of the algorithm to match the current standard of care in burn injury assessment. Given that there are few burn surgeons and burn care facilities in the United States, this technology is expected to improve the standard of burn care for burn patients with less access to specialized facilities.

Multispectral imaging burn wound tissue classification system: a comparison of test accuracies between several common machine learning algorithms

The clinical judgment of expert burn surgeons is currently the standard on which diagnostic and therapeutic decisionmaking regarding burn injuries is based. Multispectral imaging (MSI) has the potential to increase the accuracy of burn depth assessment and the intraoperative identification of viable wound bed during surgical debridement of burn injuries. A highly accurate classification model must be developed using machine-learning techniques in order to translate MSI data into clinically-relevant information. An animal burn model was developed to build an MSI training database and to study the burn tissue classification ability of several models trained via common machine-learning algorithms. The algorithms tested, from least to most complex, were: K-nearest neighbors (KNN), decision tree (DT), linear discriminant analysis (LDA), weighted linear discriminant analysis (W-LDA), quadratic discriminant analysis (QDA), ensemble linear discriminant analysis (EN-LDA), ensemble K-nearest neighbors (EN-KNN), and ensemble decision tree (EN-DT). After the ground-truth database of six tissue types (healthy skin, wound bed, blood, hyperemia, partial injury, full injury) was generated by histopathological analysis, we used 10-fold cross validation to compare the algorithms’ performances based on their accuracies in classifying data against the ground truth, and each algorithm was tested 100 times. The mean test accuracy of the algorithms were KNN 68.3%, DT 61.5%, LDA 70.5%, W-LDA 68.1%, QDA 68.9%, EN-LDA 56.8%, EN-KNN 49.7%, and EN-DT 36.5%. LDA had the highest test accuracy, reflecting the bias-variance tradeoff over the range of complexities inherent to the algorithms tested. Several algorithms were able to match the current standard in burn tissue classification, the clinical judgment of expert burn surgeons. These results will guide further development of an MSI burn tissue classification system. Given that there are few surgeons and facilities specializing in burn care, this technology may improve the standard of burn care for patients without access to specialized facilities.

Dynamic tissue phantoms and their use in assessment of a noninvasive optical plethysmography imaging device

Non-contact photoplethysmography (PPG) has been studied as a method to provide low-cost and non-invasive medical imaging for a variety of near-surface pathologies and two dimensional blood oxygenation measurements. Dynamic tissue phantoms were developed to evaluate this technology in a laboratory setting. The purpose of these phantoms was to generate a tissue model with tunable parameters including: blood vessel volume change; pulse wave frequency; and optical scattering and absorption parameters. A non-contact PPG imaging system was evaluated on this model and compared against laser Doppler imaging (LDI) and a traditional pulse oximeter. Results indicate non-contact PPG accurately identifies pulse frequency and appears to identify signals from optically dense phantoms with significantly higher detection thresholds than LDI.

Non-invasive optical imaging techniques for burn-injured tissue detection for debridement surgery

Burn debridement is a challenging technique that requires significant skill to identify regions requiring excision and appropriate excision depth. A machine learning tool is being developed in order to assist surgeons by providing a quantitative assessment of burn-injured tissue. Three noninvasive optical imaging techniques capable of distinguishing between four kinds of tissue-healthy skin, viable wound bed, deep burn, and shallow burn-during serial burn debridement in a porcine model are presented in this paper. The combination of all three techniques considerably improves the accuracy of tissue classification, from 0.42 to almost 0.77.Burn debridement is a challenging technique that requires significant skill to identify regions requiring excision and appropriate excision depth. A machine learning tool is being developed in order to assist surgeons by providing a quantitative assessment of burn-injured tissue. Three noninvasive optical imaging techniques capable of distinguishing between four kinds of tissue-healthy skin, viable wound bed, deep burn, and shallow burn-during serial burn debridement in a porcine model are presented in this paper. The combination of all three techniques considerably improves the accuracy of tissue classification, from 0.42 to almost 0.77.

Assessment of a noninvasive optical photoplethysmography imaging device with dynamic tissue phantom models

Abstract. Noncontact photoplethysmography (PPG) has been studied as a method to provide low-cost, noninvasive, two-dimensional blood oxygenation measurements and medical imaging for a variety of near-surface pathologies. To evaluate this technology in a laboratory setting, dynamic tissue phantoms were developed with tunable parameters that mimic physiologic properties of the skin, including blood vessel volume change, pulse wave frequency, and tissue scattering and absorption. Tissue phantoms were generated using an elastic tubing to represent a blood vessel where the luminal volume could be modulated with a pulsatile fluid flow. The blood was mimicked with a scattering and absorbing motility standard, and the tissue with a gelatin–lipid emulsion hydrogel. A noncontact PPG imaging system was then evaluated using the phantoms. Noncontact PPG imaging accurately identified pulse frequency, and PPG signals from these phantoms suggest that the phantoms can be used to evaluate noncontact PPG imaging systems. Such information may be valuable to the development of future PPG imaging systems.