Synthesizing global context and local precision for pathological image analysis of diverse imaging modalities

Abstract

Modern clinical practice relies on non-invasive imaging to support accurate diagnosis and treatment monitoring. Techniques such as chest X-rays (CXR), magnetic resonance imaging (MRI), computed tomography (CT), whole slide imaging (WSI), and colonoscopy images allow clinicians to examine internal body structures without the usual risks associated with surgical procedures. The wider use of digital imaging has also increased the volume of data that must be reviewed each day. Interpreting these images requires time, attention, and domain expertise. In many clinical settings, high workload can affect interpretation quality, and this issue is worsened by the shortage of trained radiologists in many regions today. Hospitals worldwide remain understaffed. Deep learning has improved automated systems’ ability to analyze medical images for tasks such as classification and segmentation. These methods learn relevant patterns directly from data and have shown good performance in controlled settings. However, the reliability in routine clinical settings remains limited. Medical images often differ in quality and appearance across patient variability and imaging devices. Variations in contrast, noise, and acquisition conditions reduce the model performance. In many cases, diseased and healthy regions look similar, which makes accurate prediction difficult. Another challenge is the lack of large, well-annotated datasets. Expert labeling is time-consuming, introduces inconsistencies, and is further complicated by class imbalance, which affects model training. In addition, many models focus on fine details and the overall structure separately, rather than combining both.