Assessment of Human Skin Burns: A Deep Transfer Learning Approach

Abstract

Purpose Accurate assessment of burns is increasingly sought due to diagnostic challenges faced with traditional visual assessment methods. While visual assessment is the most established means of evaluating burns globally, specialised dermatologists are not readily available in most locations and assessment is highly subjective. The use of other technical devices such as Laser Doppler Imaging is highly expensive while rate of occurrences is high in low-and middle-income countries. These necessitate the need for robust and cost-effective assessment techniques thereby acting as an affordable alternative to human expertise. Method In this paper, we present a technique to discriminate skin burns using deep transfer learning. This is due to deficient datasets to train a model from scratch, in which two dense and a classification layers were added to replace the existing top layers of pre-trained ResNet50 model. Results The proposed study was able to discriminate between burns and healthy skin in both ethnic subjects (Caucasians and Africans). We present an extensive analysis of the effect of using both homogeneous and heterogeneous datasets when training a machine learning algorithm. The findings show that using homogenous dataset during training process produces a biased diagnostic model towards minor racial subjects while using heterogeneous datasets produce a robust diagnostic model. Recognition accuracy of up to 97.1% and 99.3% using African and Caucasian datasets respectively were achieved. Conclusion We concluded that it is feasible to have a robust diagnostic machine learning model for burns assessment that can be deployed to remote locations faced with access to specialized burns specialists, thereby aiding in decision-making as quick as possible

Accuracy of Image-Based Automated Diagnosis in the Identification and Classification of Acute Burn Injuries. A Systematic Review

“…In fact, that study used an existing dataset from the first generation, and obtained similar accuracy results, despite using a more up-to-date methodology [18]. In the third generation, five studies split their data between training and validation prior to training [16,31,[33][34][35], while the four others used cross-validation (i.e., multiple random splitting of their dataset) as a validation technique [17,30,32,36].…”

“…This generation still requires that the images be pre-processed to feed the algorithm with image-specific features rather than images themselves. The third generation, introduced by teams from five different-and several new-countries, consists of studies from 2020 and 2021, which typically used transfer learning methodologies to increase the power of a small sample sizes of images, and which used as their inputs the whole images [16,17,[30][31][32][33][34][35][36]. Indeed, the specificity of transfer learning is such that the fine-tuning of the algorithm can be performed using only a few burn images on an already existing CNN, previously trained on a colossal dataset of unrelated objects for a different task [37].…”

“…Three studies looked at both identification and segmentation [23,24,29] (two of which are from the first generation [23,24]). Burn identification has remained a subject of interest across generations even though it can be considered of lesser clinical importance (if there is no burn, the clinicians do not need to be involved), with four studies from the latest generation studying this specifically [17,31,34,35]. Although not indicated clearly in the studies, burn segmentation could be useful for burn size estimation.…”

Burns Depth Assessment Using Deep Learning Features

“…Similarly, the study referenced in [19] proposed a binary classification of burns and healthy skin using fine-tuning approach. A pre-trained ImageNet deep learning model (i.e.…”

DeepFMD: Computational Analysis for Malaria Detection in Blood-Smear Images Using Deep-Learning Features

“…Light microscopy using blood films is currently the most popular technique for diagnosing malaria [2]. Diagnosing malaria under a microscope is conducted by applying a patient's blood drop on a glass slide, and subsequently immersing it in a staining solution to make the parasites visible.…”