Ordinal classification framework for multiclass grading of pneumoconiosis

Despite the multitude of ordinal diagnostics tasks in healthcare, deep learning classification methods usually model target classes as independent without ordinal classification. Pneumoconiosis is an occupational lung disease caused by the inhalation of mineral dust particles. Pneumoconiosis grading remains challenging considering its ordinal property and reliant on subtle variations on chest radiographs. Our work evaluates whether modeling ordinal relationships during training improves classification of pneumoconiosis profusion of small opacities (Grade 0, 1, 2, or 3) with 0 referring to normal and category 3 representing the most severe profusion. We employ both ResNet and Vision Transformer (ViT) frameworks with different loss functions to evaluate this ordinal classification task: 1) cross-entropy loss, 2) Mean-Squared Error (MSE) loss and 3) multi-task conditional loss. Training and testing was performed on a curated dataset of 708 chest radiographs with consensus readings for small opacities grade. Results show that the ResNet model with HCE loss outperforms the ResNet model with other loss functions, obtaining the highest multiclass accuracy of 71.6% (SD=0.014) on the test set. Compared to traditional cross-entropy loss, the ViT model with HCE loss achieves better performance (accuracy 57.1% with SD=0.014 on test data set). However its overall accuracy is still significantly lower than ResNet framework, potentially due to the absence of domain-specific pre-training. This preliminary study is among the first to apply deep learning to an occupational lung disease multi-classification problem and demonstrates preliminary evidence that hierarchical cross entropy loss can successfully be used to grade this disease.