Deep Segmentation for early diagnosis in Medical Imaging

KETFI., Meryem (2025) Deep Segmentation for early diagnosis in Medical Imaging. Doctoral thesis, Faculté des sciences et technologie.

Text version finale de la thèse.pdf Download (8MB)

Abstract

Every year millions around the world could be saved if they had access to faster and more accurate diagnosis of their disease. Advances in Artificial Intelligence (AI) are set to revolutionize the healthcare industry. For the year 2018, the Data Science Bowl brought together many Data Scientists from around the world to tackle one of the biggest challenges for biologists and doctors: designing image segmentation algorithms to automate the detection of nuclei in biological cells. It is no coincidence that segmentation has become a hot topic in medical image processing. Thus, researchers and doctors can understand the underlying biological processes and speed up medical diagnosis. A difficulty related to the data lies in the heterogeneity of the images. Indeed, the images are very varied, have different magnifications, different colorizations, and contain different cell types. Several strategies will therefore be required to overcome these imbalances and the size of the data set. A key characteristic of convolutional neural networks (CNNs) and Deep Learning (DL), in general, is the assumption of spatial invariance in image features. We are interested in the same patterns to be recognized in the different parts of the image. Technically, this amounts to having all the neurons of the same layer share similar weights, which considerably reduces the number of network parameters. It should be noted that this hypothesis could nevertheless limit the exploitation of very specific structures in an image, such as the geometry of an organ (face in biometrics). The work developed in this thesis brings several innovative advancements in this context. First, for classification, we used the Transfer Learning (TL) model called VGG16 after extracting parameters using the best Gabor filters to simulate retinal performance. Then, we explored a new approach to image enhancement with the extraction of global and textural features based on DL. Initially, chest X-ray and computed tomography images were preprocessed and enhanced using histogram equalization (HE). Next, global and local features were extracted using hybrid feature descriptors such as MobileNetV2 via Local Binary Pattern (LBP) models and Gabor filters. Concatenation of the best models for optimal feature extraction was employed, and DL methods for deep feature extraction and data reduction were applied for optimal classification. To validate this work, we first tested this approach on the COVID19 database (collected during the epidemic period) as well as on various types of pneumonia. For segmentation, we used the metaheuristic algorithm Particle Swarm Optimization (PSO) to improve the performance of our segmentation system. Two types of optimizations were studied: autoencoder optimization for image denoising before feeding it into the UNET model, and accuracy optimization of UNET. The results obtained are promising. All our experiments were conducted on different datasets, including COVID19, Viral Pneumonia, Breast Cancer, Skin Cancer, and a synthetic retinal database. The outcomes were satisfactory and promising, with potential for further improvement through enhanced detection and preprocessing techniques

Actions (login required)

View Item