التفاصيل البيبلوغرافية
| العنوان: |
Deep learning approaches for detecting and explaining hepatic disorders from CT scans |
| المؤلفون: |
Cuervo Noguera, Ruben |
| المساهمون: |
Universitat Politècnica de Catalunya. Departament de Ciències de la Computació, García Gasulla, Dario |
| بيانات النشر: |
Universitat Politècnica de Catalunya |
| سنة النشر: |
2024 |
| المجموعة: |
Universitat Politècnica de Catalunya, BarcelonaTech: UPCommons - Global access to UPC knowledge |
| مصطلحات موضوعية: |
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial, Deep learning, Imaging systems in medicine, Aprenentatge profund, CNN, IA Explicable, Grad-CAM, LRP, Tomografia computada, Imatge mèdica, PSVD, Cirrosi, Malaltia hepàtica, Explainable AI, CT Scan, Medical Imaging, Cirrhosis, Liver disease, Imatgeria mèdica |
| الوصف: |
Porto-Sinusoidal Vascular Disorder (PSVD) is a term that encapsulates a series of rare liver disorders characterized by specific alterations in portal vein branches which currently can only be diagnosed by performing a liver biopsy. Due to its similarity in symptomatology with cirrhosis, the most common liver disease, PSVD is often misdiagnosed. For this project, doctors from Hepatology Service of Hospital Clinic collaborated with Barcelona Supercomputing Center (BSC) to address the complexities of diagnosing PSVD by means of non-invasive methods. In order to test whether Machine Learning models could effectively learn and recognize the distinguishing patterns of PSVD, clinicians granted access to an anonymized dataset containing 201 abdominal CT Scans from four distinct groups: PSVD, cirrhosis and two control classes consisting of patients with miscellaneous liver conditions and individuals without any underlying liver disease. In this work, we analyzed the inherent biases and limitations of the dataset, revealing systematic imbalances related to sex, CT scanner type, slice thickness, and radiocontrast. To prepare the data for analysis, we developed a preprocessing pipeline that identifies and isolates the patient's abdominal cavity using bounding boxes. Subsequently, we developed a methodology for training Convolutional Neural Networks to predict individual slices, outlining the process for making predictions at the patient level. We also investigated the impact of using two distinct network architectures with varying configurations and the influence of transfer learning tailored for medical imaging. Finally, we employed Grad-CAM and LRP explainable AI techniques to visualize the learned patterns so that clinical experts can understand and interpret them, thereby enhancing their diagnostic expertise. The results indicate satisfactory performance for the task in both the train and validation sets, discriminating these pathologies through visual patterns present in CT Scans even in the presence of scarce data ... |
| نوع الوثيقة: |
master thesis |
| وصف الملف: |
application/pdf |
| اللغة: |
English |
| Relation: |
http://hdl.handle.net/2117/406465; 182806 |
| الاتاحة: |
http://hdl.handle.net/2117/406465 |
| Rights: |
Open Access |
| رقم الانضمام: |
edsbas.CD8165FF |
| قاعدة البيانات: |
BASE |