NEW STRATEGY FOR KNEE CARTILAGE ASSESSMENT THROUGH MACHINE LEARNING AND IMAGING BIOMARKERS

Osteoarthritis is a chronic, degenerative joint condition mainly affecting cartilage. This study focused on assessing knee cartilage to predict OA, a common condition causing loss of motor functions and pain in advanced stages, significantly impacting the elderly worldwide. Current clinical treatments focus on pain relief and reducing disabilities, but rely on X-rays, Kellgren-Lawrence, Ahlbäck scales, CT, and MRI, which have high variability. Many risk factors also complicate diagnosis and increase the condition's incidence. Developing new, more objective methods for evaluating cartilage is essential. This study aims to highlight which risk factors, among age, gender, and body mass index (BMI), contribute most to the manifestation of the pathology. To do this, the predictive value of the risk factors is assessed relative to the patient's condition. Furthermore, the study identifies a series of biomarkers extracted from medical images (CT and MRI) based on morphological, densitometric, and radiomic information related to knee cartilage. The goal is to develop a new cartilage assessment strategy that can identify affected individuals more quickly and effectively. In this study, 95 subjects were selected: 42 exhibited symptoms of the pathology, defined as degenerative subjects (D), and 53 healthy individuals, defined as control subjects (C). In the first phase of the study, using information on category, age, gender, weight, and height, risk factors were examined. Here, a logistic regression was used to evaluate the impact of each risk factor in distinguishing a healthy individual from an ill one. In the second part, we utilised scans from approximately 138 knees, of which 63 were degenerative (D) and 75 were healthy (C); the total exceeds the number of subjects because some subjects had both knees scanned. Subsequently, Mimics, a medical imaging software, was used to process the CT and MRI scans and extract the morphological and densitometric biomarkers of the study. From the same images, radiomic features were extracted using Python algorithms. All the extracted features were analysed using advanced statistical tools and machine learning techniques. In particular, several datasets were created with different feature combinations, which were then analysed using the potential of a support vector machine (SVM) to distinguish healthy subjects from degenerative ones based on biomarkers. Subsequently, logistic regression (LR) was applied to the SVM outputs to provide a reliability index of classification performance. The results of the first survey indicate that BMI and Age are the most significant risk factors in classifying the subject. From this, it can be inferred that these factors constitute the most relevant causes of the analysed pathology and, therefore, should be monitored with greater attention. The second analysis yields different results depending on the biomarker group considered. However, the dataset that integrates radiomic information and the dataset that combines radiomic data from CT with morphological and densitometric characteristics, after analysis with SVM, achieve weighted F1 scores above 85%. Furthermore, with the same metric, logistic regression (LR) reaches an accuracy of 98%. This promotes the features and the use of these models for early diagnosis of the pathology. This work represents a first step towards a better understanding of the pathology and the development of a targeted strategy for cartilage assessment. Indeed, the study broadens knowledge of the effects of risk factors related to the pathology and lays the groundwork for making a significant contribution to diagnosis, as this approach, which utilises features and machine learning models, can be reused in clinical settings, reducing time and costs.