Abstract
Every year, millions of children worldwide succumb to early childhood pneumonia, highlighting a significant global health concern. Despite advancements in diagnostic technology, the primary obstacle to reducing mortality rates remains the shortage of trained medical professionals. Traditional approaches to pneumonia diagnosis often rely on subjective interpretation and can be resource-intensive. In response to this challenge, deep learning models have emerged as a promising tool for automating diagnostic processes and improving healthcare outcomes. Initially, our research delved into basic classification deep learning models to identify patterns and features indicative of pneumonia from medical imaging data. These models aimed to classify images as either indicative of pneumonia or healthy lung tissue, leveraging labeled datasets to train algorithms. While these classification models demonstrated promising results, they eventually reached a performance ceiling, prompting a shift towards exploring alternative approaches. Recognizing the substantial class disparity inherent in pneumonia datasets, we then turned our attention towards anomaly detection techniques. Anomaly detection offered a novel perspective, enabling the identification of irregularities or deviations from normal lung structures that may indicate the presence of pneumonia. By leveraging unsupervised learning methods, we sought to develop models capable of detecting subtle anomalies indicative of early-stage pneumonia, thereby facilitating earlier intervention and treatment. Amidst these explorations, the concept of unsupervised encoder-decoder methods emerged as a promising avenue for pneumonia detection. These methods aimed to learn meaningful representations of lung imaging data without the need for explicit labels. One such innovation born from these investigations is the Self-Attentive Masked Encoding (SAME) method. SAME leverages the power of self-attention mechanisms within an autoencoder framework to encode and decode lung ultrasound images faithfully. By incorporating attention mechanisms, SAME can capture intricate dependencies within the input data, facilitating the extraction of meaningful features relevant to pneumonia detection.