All images were obtained from research participants who were enrolled in a diabetes education program at an academic center and who had an unknown lipohypertrophy status between July 2015 and March 2017 as part of a previous study of this condition [10]. All research participants were above 19 years of age, had a diagnosis of type 1 or type 2 diabetes mellitus, and were currently being treated with a minimum of 1 insulin injection daily or an insulin pump for at least 2 years. Participants were excluded if they were prescribed a systemic glucocorticoid, glucagon-like peptide-1 agonist, or if they had a nonlipodystrophic dermatological condition extending to the insulin injection site area. Each image was categorized as positive (lipohypertrophy present) or negative (no lipohypertrophy present) by a radiologist as per previously published criteria in a blinded fashion [10]. Ultrasound images were obtained in a blinded fashion using a portable GE LOGIQ e machine with an L8-18I-D probe (5-18 MHz; GE Healthcare).

All research participants gave written consent, and our study protocol received approval by the Human Subjects Committee of the University of British Columbia (H20-03979).

Before beginning any model training, the data were split into train, validation, and test splits of 70%, 15%, and 15%, respectively, followed by some preprocessing steps of manually removing borders from the nonannotated versions of the images. We included all different types of diabetes as 1 set and did not differentiate between patients when splitting, as the histology of these lesions has been found to be independent of the source of insulin or mode of administration [18,19]. In fact, insulin-induced lipohypertrophy does not show any histological specificity, closely resembles hypertrophic cellulite [20], and appears identical to fat nodules due to other etiologies such as corticosteroids [21] or electromagnetic fields [22]. The lesions have been shown to be due to the direct result of the hypertrophic effects of administered insulin with no evidence for a pathogenic role for the insulin antibodies found in type 1 diabetes [23].

Given the small size of the data set, image augmentation techniques were used to expand the size of the training set and improve the model’s generalizability. A variety of classic transformations [16,17] were tested, and the model’s performance on these augmented data sets were documented at this stage (Figure 1). The augmenting transformations that led to the best performance were adding random vertical and horizontal flipping, randomly changing the brightness between –0.1 to 0.1, and randomly changing the contrast between 0 and 1, each with a probability of 50%. The images in the data set varied in size from 300300 up to 460500. As a result, after the above transformations, all images were resized to a standard common denominator of 300300 pixels by cropping. An example of a transformed image is shown in Figure 1. The augmented data is then used to train a CNN model using transfer learning, a technique using pretrained models on thousands of images, which then allows for retraining of the entire network with our comparatively smaller data set. Based on our literature review, the transfer learning architectures we chose to investigate were the following: VGG16, ResNet50, DenseNet169, and InceptionV3 [24]. Each model was incorporated into our small data set, trained in separate experiments using techniques to optimize the parameters of the model to maximize its ability to learn. To compare the performance of different architectures, the team considered the accuracy and recall scores of the models when tested on our test set.

In addition, we wanted to implement object detection into our pipeline, giving users the opportunity to visually identify the location of lipohypertrophy being detected by our model. To implement object detection using a popular framework called YOLOv5 [25,26], the team created bounding boxes around the location of the lipohypertrophy masses on the positive training images using the annotated ultrasound images as a guide. Next, using the YOLOv5 framework, the YOLOv5m model was trained for 200 epochs with an image size of 320320 pixels (as this was what the Application Programming Interface allowed) and a batch size of 8.

As a proof of concept, we were able to demonstrate the ability of a supervised machine learning algorithm to detect lipohypertrophy on ultrasound images using a CNN, and we were able to deploy this algorithm though a web-based application to make accurate predictions on the presence or absence of lipohypertrophy in ultrasound images obtained at the point of care. The DenseNet transfer learning architecture outperformed the other architectures tested, suggesting this would be the most appropriate choice to automate the process of detecting and locating lipohypertrophy, a common dermatological complication of insulin injections.

Prediction of masses in ultrasound images using machine learning techniques has been an ongoing effort in clinical practice for the past few decades. To assist physicians in diagnosing disease, many scholars have implemented techniques such as regression, decision trees, Naive Bayesian classifiers, and neural networks on patients’ ultrasound imaging data [12]. Further, similar to this study, many investigators have used preprocessing techniques to extract features. In fact, Chiao et al [13] demonstrated that CNNs using ultrasound images perform better than other methods (such as radiomic models) in predicting breast cancer tumors. Another recent study showed considerable success in classifying liver masses into 1 of 5 categories with 84% accuracy, using a CNN mode [14]. To our knowledge, this is the first attempt to use CNN techniques to automate the detection of lipohypertrophy, demonstrating the considerable performance of our DenseNet model both in terms of test accuracy and recall (Table 2).

Recent research has delved into various complex image augmentation techniques to generate images [15]; we also found that traditional transformations managed to improve model performance, congruent with the results of this study. Furthermore, other studies [16,17] also confirmed that minimal transformations such as flipping the images led to higher prediction accuracy in their application. DenseNet has also proved successful in similar deep learning applications using small data sets [27], which we suspect is due to its ability to reduce the parameters in a model.

Although our project has demonstrated in principle that machine learning can be used to detect lipohypertrophy, there are some key limitations that should be addressed before it can be used in a clinical setting. Given the small size of our data set, more images need to be incorporated into the model before it can be used to direct patient care. Besides, even after the addition of new images, an auditing process should also be developed to ensure that our machine learning model does not propagate any biases that could cause harm to specific patient populations.

Previous clinical studies of lipohypertrophy have demonstrated quite a high prevalence of this condition (greater than half). More importantly, they have demonstrated a significant burden of subclinical lesions in patients with diabetes [10]. This is clinically important both due to the alterations in insulin absorption with injection proximate to a lipohypertrophic lesion [5-7] and the fact that the only treatment for this condition is avoidance [28]. Although our proof-of-concept study was limited by the fact that our model was based on a small number of images, we have successfully demonstrated the development of a model that can automatically detect lipohypertrophy in patients with diabetes. Although more work needs to be done, future studies of models developed on larger image data sets could allow for the development of a rapid, noninvasive, bedside test for subclinical lipohypertrophy that could easily be used by health care professionals unfamiliar with the use of ultrasound technology.