Back to search

University of Crete (UOC)

Artificial intelligence in bone marrow imaging: development of novel machine learning strategies for the diagnosis and classification of benign bone marrow pathology with the use of magnetic resonance imaging (MRI)

Abstract

dc:description

Bone marrow edema (BME) is a non-specific finding that can accompany a wide variety of conditions affecting the bone marrow including acute trauma, acute bone marrow edema syndromes (transient osteoporosis, regional migratory osteoporosis), chronic regional pain syndrome, avascular necrosis, infection, inflammatory arthritis, osteoarthritis at advanced stages of the disease, tendinopathies, and primary and metastatic malignancies. The imaging modality of choice for the depiction of bone marrow edema is magnetic resonance imaging (MRI) with fluid-sensitive sequences. The appearance of BME on MRI can complicate the diagnosis of diseases affecting the bone marrow, creating diagnostic dilemmas that challenge general radiologists or even specialized musculoskeletal radiologists. The most important diagnostic challenges faced in everyday radiological practice are (a) the differentiation between transient osteoporosis of the hip and avascular necrosis, (b) the differentiation between subchondral insufficiency fractures of the knee and advanced osteoarthritis presenting with BME and (c) the accurate staging of avascular necrosis. Accurate diagnosis in these cases is of utmost importance since it can change the treatment from conservative (for transient osteoporosis) to surgical (for avascular necrosis) or can determine the choice between joint preserving surgery (early stages of avascular necrosis) and total hip arthroplasty (late stages of avascular necrosis. The aim of this PhD was to leverage the power of novel image analysis methods such as radiomics and deep learning to tackle the aforementioned diagnostic dilemmas. Radiomics includes the extraction of high-dimensional data from regions of interest that can be used for the detailed characterization of lesions. Artificial intelligence (traditional machine learning or deep learning) methods can be used to either analyse radiomics data or to independently perform image recognition tasks for diagnostic purposes attempting the automation of disease detection. For the purposes of this PhD, radiomics data have been utilized for the analysis of proximal femurs with either avascular necrosis or transient osteoporosis of the hip, used to train machine learning models to distinguish between the two conditions. These models achieved excellent performance in distinguishing between the two conditions, performing equally to musculoskeletal radiologists and better than a general radiologist. In addition, in order to further automate the diagnosis between these two conditions, and to avoid bias related to the manual steps for radiomics data preparation, deep learning was used to distinguish between the two using whole images. Three convolutional neural networks (CNNs) were trained with a transfer learning methodology and finetuned with our data, in order to diagnose between transient osteoporosis and avascular necrosis. The consensus decision between the three CNNs was found to be highly accurate, performing better than two experts. Subsequently, a CNN ensemble was used to differentiate between subchondral insufficiency fractures and advanced osteoarthritis of the knee. The consensus decision of the network ensemble was compared to the diagnosis of expert radiologists. This CNN ensemble was found to be highly accurate in the differentiation between the two conditions performing better than one of the two experts. Finally, CNNs were used to distinguish between early (ARCO 1-2) and late stages of avascular necrosis (ARCO 3-4). The consensus decision of three CNNs was found to reach high performance in this diagnostic task. To further validate the model, a dataset from another country was used to assess model performance on unknown data and this validation performance was compared to the diagnosis of expert readers. Despite the performance drop in the external dataset, the CNN ensemble was still highly accurate in recognizing late AVN achieving a performance similar to the two experts. In conclusion, the work presented herein demonstrated the potential of radiomics and deep learning to assist diagnostic decisions in some of the most complicated tasks related to the presence of BME on MRI.

Degree

thesis:*
Grantor dc:publisher
University of Crete (UOC)
Year dc:date
2023

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Klontzas, Michail
  • Κλώντζας, Μιχαήλ

Subjects

dc:subject × 24

Rights

Language dc:language
eng

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/53915

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Klontzas, Michail; Κλώντζας, Μιχαήλ. Artificial intelligence in bone marrow imaging: development of novel machine learning strategies for the diagnosis and classification of benign bone marrow pathology with the use of magnetic resonance imaging (MRI). University of Crete (UOC), 2023. http://hdl.handle.net/10442/hedi/53915