Unit:
Κατεύθυνση Βιοπληροφορική-Επιστήμη Βιοϊατρικών ΔεδομένωνΠληροφορική
Author:
Chrysoula Theodora
Supervisors info:
Δρ. Θεόδωρος Δαλαμάγκας, Διευθυντής Ερευνών, Ινστιτούτο Πληροφοριακών Συστημάτων, Ερευνητικό Κέντρο Αθηνά.
Δρ. Χρήστος Δίου, Επίκουρος Καθηγητής, Τμήμα Πληροφορικής και Τηλεματικής, Χαροκόπειο Πανεπιστήμιο
Original Title:
Explainable Artificial Intelligence for Deep Learing Methods in Chest X-Ray Classification
Translated title:
Explainable Artificial Intelligence for Deep Learing Methods in Chest X-Ray Classification
Summary:
Chest X-rays are a crucial tool for detecting abnormalities with the classification and localization of these diseases under intense research. The black box nature of deep learning algorithms necessitates the development of eXplainable Artificial Intelligence (XAI) methods. This study employs the VinBigData dataset, featuring 18,000 posterior-anterior (PA) images from Hospital 108 (H108) and Hanoi Medical University Hospital (HMUH) in Vietnam.
The focus of this study is on classifying six abnormalities (‘Aortic Enlargement’, ‘Cardiomegaly’, ‘Lung Opacity’, ‘Pleural Effusion’, ‘Pleural Thickening’ and ‘Pulmonary Fibrosis’) and a ‘No-Finding’ class which represents the absence of a disease. A pretrained ResNet50 on the ImageNet dataset is used, and Grad-Cam is the chosen XAI method. Evaluation of the XAI methods involves using the Intersection Over Union (IoU) metric to assess alignment between ground truth and predicted bounding boxes. Pixel importance analysis is also used for evaluation of the XAI method by replacing crucial pixels identified by Grad-Cam, with mean values in all three channels.
The model achieves a micro F1 score of 0.81, with ‘No-Finding’ obtaining the highest F1 score (0.96). ‘Aortic Enlargement’ and ‘Cardiomegaly’ show satisfactory F1 scores (0.86 and 0.83), while ‘Lung Opacity’ and ‘Pulmonary Fibrosis’ exhibit lower values (0.55 and 0.57). Examining Grad-Cam heatmaps reveals stable behaviour and localization for ‘Aortic Enlargement’ and ‘Cardiomegaly’. However, other classes produce less reliable heatmaps, with ‘Pleural Thickening’ showing the least favourable results.
While this research provides encouraging outcomes, chest X-rays classification remains challenging, necessitating further research of XAI methods and evaluation processes.
Main subject category:
Science
Keywords:
Classification, Convolutional Neural Networks, chest X-rays, eXplainable Artificial Intelligence (XAI) methods, Grad-Cam
