A COMPARATIVE STUDY OF CONVOLUTIONAL NEURAL NETWORKS (CNN) AND INCEPTION V3 FOR FACIAL SKIN DISEASE CLASSIFICATION

Abstract

Acne and the malignant skin condition basal cell carcinoma along with other skin diseases dramatically affect global health. The promising capabilities of deep learning in dermatological classification exist only for isolated disease groups since studies exclusively analyze individual conditions without covering entire facial skin disorders. The study investigates CNN and Inception V3 models for the classification of five important conditions which include acne and actinic keratosis with basal cell carcinoma and eczema followed by rosacea. A total of 1,250 validated DermNet images were used for processing which included resizing along with normalization techniques and data enhancement methods. The Inception V3 model operates with implemented dense layers but a custom CNN model runs with an original structure of 5 convolutional and 2 dense layers. Both used Adam optimization and categorical cross-entropy loss. The test accuracy levels from Inception V3 (94%) outperformed those of CNN (93%) and the precision and recall coefficient reached 0.83 macro averages. The optimization process in CNN revealed overfitting since the training accuracy reached 88% while the test accuracy only reached 80%. Inception V3 outperformed other models by achieving F1 scores of 0.84 for rosacea and 0.86 for acne but CNN proved slightly better at identifying eczema with an F1 score of 0.81. The investigation proved that Inception V3 exhibits strong capabilities for diagnosing facial skin diseases through accurate scalable results. Despite dealing with restricted data the model kept outstanding performance levels through its implementation of transfer learning skills. Future research must increase the dataset size while conducting clinical tests to enhance universal usage. This research enables automated dermatological diagnosis to specifically identify five conditions which do not have current complete deep learning-based solutions.