Radiology-Assisted Detection of Pleural Effusion Using a Balanced Baseline Model

Abstract

Pleural effusion is a frequent and clinically relevant finding on chest radiographs in acute and critical care. An automated detection pipeline was developed using transfer learning with DenseNet121 on a strictly balanced NIH ChestX-ray14 cohort (3,955 Pleural Effusion / 3,955 No Finding; 7,910 images) partitioned patient-wise into 70% training (5,933), 10% validation (665), and 20% test (1,312). Training used restrained augmentation (RandomZoom 0.1, RandomContrast 0.1); validation and test images were not augmented. After feature-extraction pretraining, targeted fine-tuning was applied to the upper DenseNet121 blocks. Predictions were obtained from softmax class scores and metrics were computed from the resulting confusion matrices. On the test set, results were accuracy 0.824, precision 0.877, recall 0.755, specificity 0.893, and F1 0.811 (TP 497, FN 161, FP 70, TN 584). On the validation set, results were accuracy 0.824, precision 0.853, recall 0.758, specificity 0.883, and F1 0.803 (TP 238, FN 76, FP 41, TN 310). Grad-CAM analyses highlighted saliency over costophrenic recesses, basal opacities, and meniscus-like contours in correctly classified positives. These findings indicate stable discrimination with an operating profile that favors specificity, establishing a transparent and reproducible baseline for pleural-effusion detection on chest radiographs.