Improving computer aided analysis for medical diagnosis

Abstract

This thesis addresses pragmatic challenges in computer-aided diagnosis (CAD) within medical image analysis. Our research spans diverse applications, including mobile monitoring for orthodontic treatment and the integration of eye-tracking information to enhance medical image diagnosis and segmentation. These pursuits contribute to the overarching goal of making CAD more convenient and efficient.

In Chapter 3, we introduced a comprehensive framework for precise estimation of individual tooth poses in orthodontic treatment monitoring. Our method employs joint iterative optimization for camera calibration and tooth pose estimation, achieving high accuracy. Incorporating a trimming mechanism and global search strategy enhances performance. The introduced relative pose bias provides robust evaluation, surpassing precision requirements for dental applications, using the convenience of mobile cameras.

Chapter 4 presents a pioneering framework for medical image segmentation using gaze information as weak supervision. Addressing scale differences between segmentation classes, our method outperforms existing models and effectively handles limited clinical data. The integrated gaze data collection system proves unobtrusive, offering valuable information without disrupting clinical workflows.

In Chapter 5, we propose a gaze-enhanced medical image semi-supervised learning framework for classification, mitigating limited labeled data challenges. Leveraging unlabeled medical images and gaze information, our approach, combined with self-supervised pre-training, outperforms state-of-the-art methods. Overcoming data scarcity, it showcases adaptability in addressing constraints posed by limited labeled data in medical imaging datasets.