DEEP LEARNING ALGORITHMS FOR BIOMEDICAL IMAGE SEGMENTATION IN LOW-DATA SCENARIOS

Abstract

Automatic segmentation via deep learning plays a major role in biomedical imaging, enhancing diagnostics by dividing images into regions of interest. This procedure helps medical experts understand disease characteristics, lesion sizes, and other crucial details. Despite its potential, deep learning-based automatic segmentation often relies on large annotated data to accurately predict lesions and other critical regions. Among imaging modalities, ultrasound, widely used for its accessibility, real-time capabilities, and effectiveness in detecting lesions, remains inadequately investigated due to the inherent challenges in medical imaging, such as data availability and privacy concerns. This work identifies key research gaps in ultrasound imaging segmentation to address these challenges and contributes to advancements in this critical area.

This dissertation focuses on three key areas for advancing ultrasound image segmentation and improving biomedical image analysis. First, it aims to improve supervised learning-based architectures for tumor segmentation, particularly U-Net models, which, despite their success in biomedical segmentation, often lack reliability for clinical use, especially when tested on out-of-dataset samples. Second, it addresses the challenges posed by limited annotated ultrasound data, which restricts the performance of supervised models. Finally, it addresses the scarcity of ultrasound datasets paired with corresponding masks, a significant issue caused by data privacy concerns, the lack of datasets from various countries, and the high costs of expert-level annotations.

This dissertation introduces an improved supervised model based on a refined U-Net architecture incorporating ReSidual U-blocks (RSU) and Attention Gates to address segmentation challenges in scenarios with limited data. These enhancements improve the model’s ability to capture critical features and long-range dependencies, improving lesion segmentation performance in ultrasound images. Building on this, we integrate a Denoising Diffusion Probabilistic Model (DDPM) with the RSU architecture to create a deeper network capable of handling the high variability and noise in ultrasound datasets. This combination enhances segmentation mask accuracy and addresses challenges posed by samples with diverse characteristics, such as size and shape variations of regions of interest.

Next, we improve data augmentation by enhancing the Mixup technique to address limited data scenarios in image segmentation. Using K-means clustering, ultrasound images are grouped into clusters of similar samples, and Mixup applies within clusters. This approach has the potential to reduce randomness, avoid mixing unrelated regions like tumors and dark backgrounds, and ensure more effective augmentation. It also diversifies the dataset by generating new samples and masks, mitigating data scarcity. Building on this contribution, we extend the application of Cluster Mixup to unsupervised segmentation. The goal is to leverage unlabeled ultrasound images by augmenting healthy samples with Cluster Mixup, followed by unsupervised learning to detect suspected tumors. This approach could show the potential to qualitatively and quantitatively improve the segmentation of regions of interest and enhance diagnostic capabilities.

Additionally, we build upon Cluster Mixup by proposing a variant of D-DDPM, a diffusion-based model, to learn the distributions of combined images and masks, enabling the simultaneous and joint generation of synthetic images and annotations. This technique expands the dataset with a large number of image-mask pairs. We involve medical experts in evaluating the synthetic dataset, ensuring the selection of relevant samples, and improving dataset quality. Statistical analysis obtained from medical experts shows the reliability of our approach and its potential application to real-world problems.