Morphology-Constrained Prototype Learning for Explainable Pulmonary Function Screening

Abstract

Pulmonary function screening is a critical procedure for the early detection and management of chronic respiratory diseases, including chronic obstructive pulmonary disease and asthma. While deep learning models have achieved remarkable success in various medical diagnostic tasks, their deployment in clinical settings is often hindered by their opaque, black-box nature. This paper presents a novel approach termed Morphology-Constrained Prototype Learning for explainable pulmonary function screening. By integrating domain-specific physiological knowledge into the latent space of a neural network, the proposed methodology constrains learned prototypes to reflect clinically meaningful morphologies derived from spirometry flow-volume loops. This approach not only maintains high diagnostic accuracy but also provides case-based interpretability, allowing clinicians to understand the model predictions by comparing patient data to learned morphological prototypes. The methodology leverages a convolutional encoder coupled with a specialized prototype layer, optimized through a composite loss function that includes classification, clustering, and morphological alignment terms. Comprehensive evaluations on large-scale clinical datasets demonstrate that the proposed model achieves competitive performance against traditional black-box architectures while offering unprecedented transparency. The results underscore the potential of combining data-driven feature extraction with physiological constraints to bridge the gap between artificial intelligence and clinical trust.