Contribution to the categorization of chest x ray radiographic

Abstract

Chest X-ray radiography is the most widely performed imaging modality worldwide, valued for its accessibility, affordability, and diagnostic relevance in clinical practice. Beyond its medical applications, chest X-rays also capture skeletal structures that are distinctive and relatively stable over time, which opens the possibility of using them as a novel biometric modality. This potential is particularly relevant in forensic and postmortem contexts, where traditional biometric traits such as fingerprints, iris, and facial features are often degraded or unavailable due to decomposition or trauma. This thesis investigates the feasibility of chest X-rays as a biometric trait and proposes a series of contributions toward developing a robust identification system. The first contribution establishes a proof of concept by applying deep learning architectures, specifically Siamese networks trained with triplet loss, to chest radiographs. Several pre-trained convolutional neural networks (VGG16, VGG19, ResNet50/101, and Dense-Net) were evaluated as backbone encoders to extract discriminative features for person identification. Building on this foundation, the second contribution introduces the Self-Residual Attention Network (SRAN), a modified backbone designed to improve feature representation. By integrating spatial and channel attention mechanisms with residual connections, SRAN enables the system to emphasize critical anatomical regions while suppressing irrelevant information, thereby enhancing robustness and discrimination. The third contribution addresses the limitations posed by soft tissues, which are susceptible to pathological variations and postmortem degradation. To overcome this, an Attention U-Net segmentation model was employed to isolate skeletal structures such as ribs, clavicles, sternum, and vertebrae. This step ensures that the biometric system relies exclusively on stable anatomical features.

The final contribution integrates segmentation and attention-based feature learning into a complete skeletal- based chest X-ray biometric framework. This system generates compact embeddings that allow both

verification (1:1) and identification (1:N), making it suitable for forensic scenarios where only skeletal information may be available. The proposed approach was evaluated using large-scale public datasets, including NIH ChestX-ray14 for training and CheXpert for testing. The results demonstrate that chest X-rays, particularly when focused on skeletal structures, offer distinctive and reliable identity cues. By introducing chest radiographs as a new biometric modality, this work contributes to both the scientific field of AI-driven biometrics and the practical domain of forensic science, offering a pathway toward robust postmortem and medico-legal identification.