Severity Assessment of Psoriatic Plaques Using Deep CNN Based Ordinal Classification

Development of computer-aided diagnosis (CAD) tool for severity assessment of psoriatic plaques is important to assist the dermatologists to overcome the human limitation. In this paper, a pioneering attempt is made to build a Convolutional Neural Network (CNN) model to classify a skin image with respect to its severity class. However, the commonly used loss functions like categorical cross entropy and mean square error ignores the underlying ordinal class relationships (distance between predicted and actual class) which are important for the present problem. In this paper, the Earth Mover’s Distance based loss function is proposed for training CNN since it takes into account the corresponding ordinal class relationships. Separate CNNs are trained for severity scoring corresponding to three plaque characteristics- erythema (redness), scaling (silveryness) and induration (elevation). Mean accuracy (MA), mean absolute error (MAE) and Kendall’s \(\tau _b\) are used for performance evaluation. The experimental result shows that the proposed ordinal classification technique outperforms the traditional approaches.