According to 2 new datasets from Jilin University, a new proposal for grading acne via set size has been shown to perform better than previous cutting-edge methods.
Severe or inflammatory acne can lead to significant scarring in affected patients, highlighting the importance of acne classification in diagnosing the skin condition. However, grading acne is often labor intensive and can lead to mistakes made by dermatologists.
As such, investigators led by Shuai Liu, MD, College of Computing and Technology, Jilin University, felt it was imperative to develop thoughtful and automatic diagnostic methods to classify acne.
With their study, Liu and colleagues proposed a novel ensemble classification framework for classifying acne severity, which relates to an ensemble pruning strategy designed to reduce the computational complexity of the classification model. together formed.
Additionally, the team used the base model prediction results as a new feature set and used the classifier to cluster the base model results.
Liu and colleagues used the ACNE04 dataset, which contains 1457 images, in evaluating the proposed algorithm, AcneGrader, for acne detection and grading. This dataset annotated local lesions and overall acne severity using the Hayashi criterion determined by professional dermatologists.
Investigators used 80% of the randomly retrieved samples to train the model they developed, while the remaining 20% were used to test the model. Specifically, the dataset contained 3297 images, of which 2637 were used for training and 660 for testing. A quintuple stratified cross-validation strategy was used to evaluate the prediction algorithms, and each of the quintuples was iteratively used in the test dataset.
Notably, a dermoscopic skin cancer image dataset was used to further verify the effectiveness of the model.
The study formulated the classification problem of acne as a 4-class classification problem, and acne was classified as mild, moderate, severe, and very severe.
Redundancy patterns were then removed by a feature selection algorithm, followed by the team integrating all base patterns through classifiers. Finally, the ensemble pruning algorithm has been proposed to prune deep learning models.
Additional training samples did not appear to lead to improved prediction performance, and 80% of the training dataset generated the best models.
However, the researchers noted that the pruning algorithms improve the overall learning models, overall, and the Kappa statistics achieved the best model performance using 22 base models, which were chosen as default model.
The experimental data indicated that the ensemble pruning framework resulted in an 85.82% prediction accuracy on the acne dataset, which the investigators said was better than existing studies.
“If our framework is used in an environment with limited computing resources, such as mobile devices, the ensemble model pruned by the error pruning strategy can be considered,” the team wrote.
The study, “AcneGrader: An Ensemble Pruning of Basic Deep Learning Models for Acne Grading,” was published online in Skin research and technology.