Explainable deep learning-based multi-class classification of inflammatory dermatoses from histopathological images

Abstract Dermatitis herpetiformis, discoid lupus erythematosus, eczema, and lupus vulgaris are inflammatory and granulomatous skin diseases with overlapping histopathological features, often leading to diagnostic uncertainty. Existing diagnosis relies on expert histopathological interpretation, while machine learning-based studies addressing multi-class classification using histopathological images remain limited, particularly in resource-constrained settings. Therefore, this study aimed to develop and evaluate deep learning models for classification of four dermatological conditions using histopathological images. A balanced dataset comprising 96 original images per class was collected from Khyber Medical College. Offline augmentation using Albumentations was applied exclusively to the training set, resulting in 486 images per class, including original and augmented images. Four pre-trained deep learning models ResNet50, MobileNetV2, VGG19, and EfficientNet-B0 were fine-tuned using transfer learning for four-class classification. Models were evaluated using an approximately 85:15 patient-level training-validation split, with 60 original images from 20 validation patients forming the validation cohort. Performance was assessed using accuracy, precision, recall, F1-score, AUROC, AUPRC, and confusion matrices. ResNet50 achieved an accuracy of 96.67% and a micro-average AUROC of 0.996, with two misclassifications among the 60 validation images. EfficientNet-B0 achieved 95.00% accuracy with a micro-average AUROC of 0.990, followed by MobileNetV2 (85.00%) and VGG19 (78.33%). For ResNet50, the confusion matrix showed 15/15 correct classifications for eczema and lupus vulgaris, while DLE and dermatitis herpetiformis each had one misclassified image. Grad-CAM + + provided qualitative visualization of image regions associated with model predictions. A web-based prototype was developed as an AI-assisted research prototype for supportive image classification. The findings demonstrate the feasibility of deep learning for four-class histopathological image classification, while the small, single-center dataset and absence of external validation limit conclusions regarding generalizability and clinical applicability. Future studies should evaluate larger multi-center datasets and incorporate additional clinical information.