Leveraging Channel Slicing and Overcomplete Network for Scaliness Estimation in Psoriasis Lesion

Accurate assessment of scaliness in psoriasis lesions is crucial for understanding disease severity and tailoring appropriate treatment strategies. Traditional methods rely on subjective visual evaluations, leading to inconsistent and error-prone results. To address this, we present a novel deep learning-based approach utilizing Overcomplete-Channel Slicing (OCS) integrated with a 3D ConvNeXt backbone for enhanced scaliness severity classification in psoriasis lesions. The proposed framework isolates fine-grained color and texture information from each RGB channel through channel slicing. Then, it applies an overcomplete representation, which projects the sliced features into a higher-dimensional space by reassembling the spatial information into the channel dimension, enabling the capture of subtle scaliness variations across the lesion. Leveraging 3D ConvNeXt’s advanced feature extraction capabilities, our model effectively emphasizes surface roughness and underlying lesion structure from 2D images, which are critical for distinguishing between absent(0), mild(1), moderate(2), severe(3), and very severe(4) scaliness. Extensive experiments on a clinically validated psoriasis dataset demonstrate that our 3D ConvNeXt-OCS framework achieves an outstanding accuracy of 0.9867, significantly outperforming various backbone networks. These results highlight the potential of our approach as an automated, objective tool for monitoring psoriasis progression and optimizing therapeutic interventions.