OPTIMIZATION OF DEEP LEARNING BASED LIGHT WEIGHT SKIN DISEASE CLASSIFICATION MODEL USING TRANSFER LEARNING AND MODEL QUANTIZATION FOR RESOURCE-CONSTRAINED ENVIRONMENTS

Abstract

Skin disease classification using dermoscopic images is one of challenging diagnostic process because of its visual similarities among various disease conditions, imbalance classes of diseased images, and the unavailability of annotated data. Many of Deep learning models have demonstrated encouraging detection performance. However, the most of Deep Learning Models requires high computational and memory requirements, that makes their deployment difficult in resource-constrained environments remains. In this study, we investigate an efficient convolutional neural network-based framework for multi-class skin disease classification using MobileNetV2. This technique is specially used for the modelling of systems for resource-constrained environments. The proposed technique will progressively enhance the performance of the model in three stages. The modelling will start from the baseline in initial stage and then considering data augmentation and unfreezing the layers. In middle stage partial fine-tuning, and class weighting to address class imbalance is considered for further improvement. Additionally, post-training quantization is used to reduce model size and enhance deployment potential with a little trade-off in performance. Experimental results show the improvements are moderately enhanced for performance macros across our three training stages. Quantization were achieved for a significant memory reduction while maintaining competitive performance of the model. This study highlights the suitability of the proposed approach for mobile based clinical applications