Summary:

Metastases are a group of abnormal cells that develop outside the confines of the original organ and spread to other organs. In particular, bone metastases originate from an organ in the body, such as the breast, lung or prostate, and spread to the bone. Although this disease was discovered over a century ago, it is still not well defined and existing treatments are not very effective, probably because it is difficult and time-consuming to detect. To help doctors, new machine learning techniques offer a solution for improving overall accuracy. This thesis aims to help radiologists systematically detect bone metastases using machine learning algorithms. However, the discovery of methodological biases in studies on the diagnosis of bone metastases and the lack of consensus on the interpretability of machine learning have shifted the focus of this thesis. It now focuses mainly on data collection and on solving the problems of validation and interpretability of machine learning. In order to properly assess the ability of machine learning to detect bone metastases, three experimental studies were conducted. The first proposed a new segmentation approach supported by an attention mechanism for locating bone lesions. The second is a study of machine learning methods for identifying cases of bone metastases. Finally, the last study highlights the lack of robustness of classification using machine learning methods and proposes a method for improving accuracy based on both the CNN and Transformer approaches. The experimental results of this thesis are evaluated on the BM-Seg dataset that we introduced, which is the first reference dataset for the segmentation and classification of bone metastases using CT scans. This open source database was used to improve the reproducibility of machine learning experiments. The results of the various preliminary studies are encouraging and promising.

Abstract:

Metastases are a group of abnormal cells that develop outside the confines of the original organ and spread to other organs. In particular, bone metastases originate from an organ in the body, such as the breast, lung or prostate, and spread to the bone. Although this disease was discovered over a century ago, it is still not well defined and existing treatments are not very effective, probably because it is difficult and time-consuming to detect. To help doctors, new machine learning techniques offer a solution for improving overall accuracy. This thesis aims to help radiologists systematically detect bone metastases using machine learning algorithms. However, the discovery of methodological biases in studies on the diagnosis of bone metastases and the lack of consensus on the interpretability of machine learning have shifted the focus of this thesis. It now focuses mainly on data collection and on solving the problems of validation and interpretability of machine learning. In order to properly assess the ability of machine learning to detect bone metastases, three experimental studies were conducted. The first proposed a new segmentation approach supported by an attention mechanism for locating bone lesions. The second is a study of machine learning methods for identifying cases of bone metastases. Finally, the last study highlights the lack of robustness of classification using machine learning methods and proposes a method for improving accuracy based on both the CNN and Transformer approaches. The experimental results of this thesis are evaluated on the BM-Seg dataset that we introduced, which is the first reference dataset for the segmentation and classification of bone metastases using CT scans. This open source database was used to improve the reproducibility of machine learning experiments. The results of the various preliminary studies are encouraging and promising.