JMIR Publications

ABSTRACT

Background:

The quality of the images received for Teledermatology evaluation is often suboptimal, with up to 50% of patients providing images that are poorly lit, off-center, or blurry. To ensure a similar level of care to in-person consultations, high-quality images are essential.

Objective:

The aim of this project is to develop an image quality analysis (IQA) tool to assess patient and primary care physicians derived images using a deep learning model, leveraging multiple instance learning and ordinal regression for model predictions.

Methods:

The dataset used for this study was acquired from patient images submitted to the Department of Dermatology at Duke University between August 21, 2018 and December 31, 2019, and PCP derived images between March 01, 2021 and June 30, 2022. Seven dermatology faculties at the level of professor, associate professor and assistant professor evaluated 400 images each, two dermatology residents evaluated 400 images, assuring that each image had four different quality labels. We used a pre-trained model VGG16 architecture, then fine-tuned by updating weights based on the input data. The images are taken with cell phones (Patients) or cameras (PCP) in RGB scale with resolution being 76 pixels per inch for both height and width and average pixel size of the image with standard deviation being 2840×2793 ± 986×983 (1471 inch2 ± 707 inch2). The optimal threshold was determined by Youden’s index which represents the best trade-off between sensitivity and specificity and balance the number of true positives and true negatives in the classification results. Once the model predicts the rank, the ordinal labels are transformed to binary labels by using a majority vote as the goal is to distinguish between two distinct categories (good vs. bad quality) and not predict quality as a continuous variable.

Results:

From the Youden’s index, we achieved a positive predicted value (PPV) of 0.906 implying that the model will predict 90% of the good quality images as such, while 10% of the bad quality images are predicted as good quality to enhance clinical utility with AUC for the test set at 0.885 (95%CI 0.838 - 0.933), sensitivity, specificity, and negative predictive value (NPV) at 0.829, 0.784, and 0.645 respectively. Further evaluation on the independent validation consisting of 300 images from patients and 150 images from physicians demonstrated AUC 0.864 (95%CI: 0.818 - 0.909) and AUC 0.902 (95%CI: 0.85 - 0.95) respectively. The sensitivity, specificity, PPV and NPV for the 300 images were 0.827, 0.800, 0.959 and 0.450 respectively.

Conclusions:

In the current work, we demonstrate a practical approach to improve the image quality for clinical decision making. While patients and PCPs may have to capture additional images (due to lower NPV), this is offset by the reduced workload and improved efficiency of clinical teams due to receiving higher quality images. Additional images can also be useful if all images (good or poor) are transmitted to the medical charts. Future works need to focus on real time clinical validation of our results.