NUST Institutions Library Catalogue

Deep learning has improved medical image analysis, yet chest X-ray systems still face
barriers in routine clinical use. Two recurring issues are severe class imbalance in clinical datasets and limited clarity about how models reach decisions, which can reduce
clinician condence. This thesis presents a neural network for multi-disease detection
in chest radiographs that reects radiology practice by organizing ndings into four
clinical groups: infectious/inammatory, structural lung changes, uid/cardiac conditions, and mass-related signs. Each group is handled by a dedicated attention module,
and a routing unit emphasizes the experts most relevant to each image. To address
imbalance, an adaptive weighting scheme monitors class-level performance during training and increases focus on classes that lag while avoiding excessive emphasis on those
already learned well. Evaluation on 3,261 X-rays across 17 diseases achieves 92.03%
accuracy and a macro AUC of 0.9729, with a trainingvalidation gap below 4%; 16 of 17
classes reach F1-scores above 80%, including several minority categories. Grad-CAM
visualizations highlight clinically plausible regions (e.g., heart border for cardiomegaly,
costophrenic angles for eusion, and lung elds for pneumonia), and a web interface
demonstrates real-time inference, bilingual output, and automated reporting. The results indicate that clinically structured attention combined with adaptive training can
provide accurate predictions with explanations that are easier to review at the point
of care.