FML loss improves the classification performance of deep learning models for uncertainty pathologies in chest radiographs

Chest radiography is the most commonly used imaging method globally. Chest radiography is a prevalent imaging technique worldwide, and developing effective computer-aided diagnosis (CAD) systems for interpreting chest X-rays necessitates the proper management of uncertainty pathologies (UP). Our study addresses the underexplored area of distinguishing between certainty pathologies (CP) and UP in CAD systems. We employed deep learning models (DLMs) to extract and compare features from CP and UP pathology samples. Furthermore, we assessed the efficacy of various DLMs with different loss functions in modeling UP feature spaces and introduced a focal misclassification label (FML) loss method to enhance the extraction of ambiguous features from UP. Our findings reveal substantial differences in the feature spaces of CP and UP, with distinct boundaries observable between the features of deterministic and uncertain pathologies across nine different conditions. We developed a novel approach, feature boundary aggregation (FBA), for processing UP during data preprocessing, treating samples within the UP feature boundary as distinct classes. The FML loss outperformed standard loss functions in capturing the nuances of UP features. This research provides valuable insights for advancing CAD systems, offering a framework for determining the feasibility of diagnosing chest anomalies from chest X-rays without additional tests. DLMs equipped with adaptable feature extraction and loss functions that emphasize misclassified labels are better suited for modeling UP in chest X-ray radiographs.