Αξιολόγηση μεθόδων ταξινόμησης και ερμηνευσιμότητας σε ακτινογραφίες θώρακος
Evaluating classification and interpretability on chest X-ray images
View/
Keywords
Νευρωνικά δίκτυαAbstract
Automatic detection of thoracic diseases in chest radiographs using deep learning is
already established as a useful methodology for improving diagnostic efficiency.
This paper presents multi-label classification of chest X-ray images with four
convolutional neural network (CNN) models – DenseNet121, Xception, EfficientNet,
and an Ensemble of these networks – trained and evaluated on the NIH ChestXray14 dataset (~112,000 images, 14 disease categories). The focus is twofold:
maximizing classification performance and enhancing model transparency through
interpretability techniques. The best results were achieved with an Ensemble model,
which outperformed individual networks in average AUC. DenseNet and Xception,
both deep high-capacity CNNs, slighthly outperforming EfficientNet in detecting
pathologies. To address the “black-box” nature of CNNs, Gradient-weighted Class
Activation Mapping (Grad-CAM) and SHapley Additive ExPlanations (SHAP) were
applied to visualize model decision regions. Quantitatively, explanation fidelity was
evaluated using Area Over the Perturbation Curve (AOPC), which reflects how much
the model’s output drops when important image regions are removed. The findings
confirm that the ensemble classifier accurately detect multiple co-occurring thoracic
diseases (average AUC ~0.80–0.82) while the use of interpretability methods
provides insights into the features driving each prediction.
Postgraduate Studies Programme
Ψηφιακά Συστήματα και ΥπηρεσίεςDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνNumber of pages
92Language
GreekCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα