Optimized Deep Neural Network for High-Precision Psoriasis Classification from Dermoscopic Images

Abstract

Accurate classification of psoriasis is critical in dermatological diagnostics due to the disease’s diverse clinical presentations and varying severity levels. With numerous subtypes and their visual similarities to other dermatological conditions, precise diagnosis typically requires expert medical knowledge. Early and accurate identification of psoriasis subtypes is essential for initiating timely treatment. This study introduces a novel hybrid deep learning architecture that integrates EfficientNet wit h Long Short-Term Memory (LSTM) networks for the automated classification of psoriasis from dermoscopic images. The proposed model is designed to simultaneously capture spatial features through EfficientNet and temporal or sequential patterns via LSTM units, thereby improving classification performance. The models are trained and tested on a publicly benchmark dataset comprising 7 distinct classes using the publically available benchmark dataset by Dermnet and BFL-NTU. Experimental results demonstrate that the proposed architecture significantly outperforms the baseline models such as VGG16 and ResNet50, with superior accuracy 89.7% and robust performance across metrics like recall, F1-score with 88%, and Region of Convergence (ROC) of 97%. This compact design with low trainable parameters reduces the computational time and memory makes the model well-suited for deployment for portable devices and enabling real-time mobile-based dermatological assessments.