Abstract
Histopathology image analysis is crucial for cancer diagnosis, prognosis, and medical research, aiding in the identification of cellular abnormalities and guiding treatment decisions. The integration of Artificial Intelligence (AI) with digital pathology is transforming this field, offering immense potential to improve patient care and advance medical knowledge. However, AI models rely on large, annotated datasets, which are often difficult to obtain due to time, resource constraints, and privacy concerns. These challenges are further amplified for rare diseases and diverse data representation. To address these limitations, my research focuses on histopathology image synthesis using conditional Generative Adversarial Networks (cGANs).
A key challenge in histopathology image synthesis is generating high-quality images with clear nuclei contours, especially for overlapping or touching nuclei. Existing cGAN-based approaches struggle with this, limiting their applicability in medical diagnostics. To overcome these limitations, I propose novel cGAN architectures enhanced with advanced attention mechanisms and tailored loss functions to improve image quality and segmentation performance.
First, I introduce SharpGAN, which employs a sharpness loss function and a normalized nucleus distance map to enhance nuclei contour contrast. This method is evaluated on two public datasets using image quality metrics and segmentation tasks.
Second, I introduce Enhanced SharpGAN. I develop a nuclei topology and contour regularization approach, integrating skeleton maps to separate touching nuclei and introducing novel contour regularization terms. This method improves segmentation performance, achieving state-of-the-art results on the Triple Negative Breast Cancer (TNBC) dataset.
Lastly, I propose an RGB contour-regularized cGAN with an attention block, incorporating topology loss to enhance nuclei homogeneity. This approach further refines synthetic image quality and segmentation accuracy, evaluated on the CoNSep dataset.
My research advances histopathology image generation, paving the way for more realistic and clinically useful synthetic datasets that can enhance AI-driven medical diagnostics.