Thesis Title: Linear and Nonlinear MIL Approaches for Medical Image Classification
Supervised by: Prof. Antonio Fuduli and Dr. Matteo Avolio
Abstract: Medical image classification is vital in modern healthcare, assisting physicians in accurate diagnosis and treatment planning. As the volume and complexity of medical imaging data increase, the demand for robust and efficient classification methods is increasingly pronounced. In this context, Multiple Instance Learning (MIL) has emerged as a powerful paradigm, providing a unique approach to address the challenges inherent in medical image classification. This study explores the complexities of medical image classification, focusing specifically on linear and non-linear multiple-instance learning approaches. Multiple instance learning, a subset of machine learning, is particularly suitable for situations where labeled training data are inadequate to obtain. This accommodates the inherent ambiguity of medical imaging, where not all regions may exhibit the pathology of interest.
In this thesis, we focused on two types of medical imaging, eye fundus, and dermoscopy, and segmented them based on their characteristics. We also applied the binary MIL method to determine positive and negative sets. We segment the blood vessels and optic disc from the eye fundus image. On the other hand, in dermoscopy (a.k.a melanoma) images, we segment the lesions. Typical image processing techniques and algorithms have been used during the segmentation process. Compared with the standard image, the segmented image provides better classification in terms of accuracy and sensitivity. Since the results seem promising, a MIL technique may be the basis for more State-of-the-art equipment to assist physicians in the medical field.
Thesis Title: Application of Optical Flow for Mobile Robot
Supervised by: Monirul Hasan
Abstract: An autonomous robot that can localize itself and navigate around can be used in various real life applications. In this research we have built a low cost, low power consuming, vision based autonomous robot. Our two wheel robot platform houses a single board computer Raspberry Pi for its computation and Arduino uno to drive the motors as instructed. The visual sensory input is coming through a webcam mounted on the robot platform itself. When the robot is in motion, we compute optical flow from successive frame capture of the webcam using OpenCV’s robust API. This allows us to navigate the robot in an unknown environment avoiding obstacles.