GravityNet for small lesion detection and classification

This thesis introduces GravityNet, a novel one-stage end-to-end deep learning framework developed specifically for the detection and classification of small lesions in medical images. Small lesions are localized abnormalities within the body, characterized by their limited spatial extent relative to the surrounding organ. Detecting and accurately characterizing these lesions are critical for the early diagnosis and treatment of various diseases. However, their low contrast and subtle appearance pose significant challenges in medical imaging. GravityNet addresses these challenges using gravity points, a novel pixel-based anchor that dynamically moves toward lesions, thereby improving detection accuracy. The framework was validated across diverse medical imaging scenarios: - Microcalcification detection in mammography: these tiny calcium deposits can serve as early indicators of breast cancer and require high sensitivity to small, low-contrast features within dense tissue. - Microaneurysm detection in retinal imaging: critical for diagnosing Diabetic Retinopathy, this task involves accurately identifying subtle vascular abnormalities. - Large Vessel Occlusion in Computed Tomography Angiography: in brain imaging, this is vital for diagnosing ischemic stroke, where rapid and precise identification of blocked vessels is crucial. - Computational Cytology: applied to whole slide images, this involves detecting and analyzing cell nuclei to identify abnormal growth patterns, key indicators of malignancies such as cervical and oral cancers. The GravityNet framework demonstrates significant potential as a versatile and powerful tool for small lesion detection across a broad range of medical imaging applications. The successful implementation in these diverse domains underscores its adaptability to various imaging modalities and highlights its ability to enhance diagnostic accuracy, supporting timely and informed clinical decision-making.