PNEUMOTHORAX SEGMENTATION ON CHEST X-RAYS: PROGRESSIVE FINE-TUNING AND THRESHOLD-BASED PREDICTION REFINEMENT

Abstract

Pneumothorax segmentation on chest radiographs remains a critical challenge for computer-aided triage, as small pleural separations are difficult to reliably localize. This study presents a segmentation method based on a U-Net-style deep neural network with a pretrained convolutional encoder, progressively fine-tuned on the SIIM-ACR Pneumothorax dataset. Training utilizes a two-stage resolution approach – starting at 512×512 for stable localization, then refining at 1024×1024 to recover thin pleural boundaries. Optimization incorporates staged encoder unfreezing, cyclic learning-rate scheduling, and a loss curriculum transitioning from weighted binary cross-entropy to soft Dice and symmetric Lovasz objectives. The experimental analysis is conducted in three stages. First, three encoder backbones (ResNet34, EfficientNet, and SE-ResNeXt-50) are compared. Second, the best-performing backbone is evaluated across confidence thresholds to determine the optimal threshold-only operating point. Third, the model is assessed under joint confidence-threshold and minimummask- size control, where excessively small predicted masks are reset to empty. This structure reflects the SIIM-ACR evaluation protocol, where correct empty-mask predictions on negative studies heavily impact the overall mean Dice score. Results identify SE-ResNeXt-50 as the strongest backbone. While a 0.70 confidence threshold maximizes threshold-only performance, the strongest overall result is achieved by combining a 0.55 confidence threshold with a minimum positive-pixel threshold of 2000. These findings demonstrate that explicit operating-point design – converting continuous probability maps into binary clinical decisions – is essential for pneumothorax segmentation and should be reported alongside architecture and training methodologies.