A Novel EfficientMobileDenseNet for Monkeypox Detection and Classification

Abstract

Monkeypox cases have become a global concern. The current method for detecting Monkeypox in humans is through Polymerase Chain Reaction (PCR) tests. However, the results obtained from this test are not accurate as they can become uncertain due to the virus remaining in the blood for a short period, and the PCR test also requires a relatively longer waiting time. DCNN models are an approach in machine learning where a model that has been previously trained on one task or domain is used as a foundation to solve another, similar, or related task. Typically, complex DCNN models produce good performance as well. However, complex transfer learning models have millions to hundreds of millions of parameters, which directly impacts longer training processes and higher computational burdens. This study presents the truncating layer enhances computational efficiency by removing unnecessary deep layers, while the merging layer creates a robust multi-feature fusion technique. These strategies fuse multiple pretrained models (EfficientNetB0, MobileNetV2, and DenseNet121) into a single powerful classifier referred to as EfficientMobileDenseNet, applied to two skin datasets: MSLDv2 (monkeypox skin lesion dataset v2), and SLD (skin lesion dataset). The proposed method EfficientMobileDenseNet achieves an accuracy of 0.85, improving over individual models, It also balances precision (0.87) and recall (0.81) effectively, Specificity (0.96) and AUC (0.97) indicate strong discrimination ability. Analysis of SLD Dataset, The proposed method achieves an accuracy of 0.83, similar to MobileNetV2 but with better recall (0.85) and higher F1-score (0.83). It maintains a high AUC (0.97), demonstrating strong classification performance and training time (579.81s) is the lowest, indicating high efficiency with parameter size of 2,255,835, which is smaller compared to the parameters of other state-of-the-art and pretrained DCNN models.