[Paper Review] Pneumothorax Segmentation: Deep Learning Image Segmentation to predict Pneumothorax

Karan Jakhar, Avneet Kaur+1 more|arXiv (Cornell University)|Dec 16, 2019

COVID-19 diagnosis using AI25 references22 citations

TL;DR

This paper proposes a U-Net-based deep learning model with a ResNet backbone for semantic segmentation of pneumothorax in chest X-ray images, achieving a Dice coefficient of 84.3% and IoU of 82.6%, enabling accurate, real-time detection and localization of pneumothorax to assist radiologists in timely diagnosis and treatment decisions.

ABSTRACT

Computer vision has shown promising results in medical image processing. Pneumothorax is a deadly condition and if not diagnosed and treated at time then it causes death. It can be diagnosed with chest X-ray images. We need an expert and experienced radiologist to predict whether a person is suffering from pneumothorax or not by looking at the chest X-ray images. Everyone does not have access to such a facility. Moreover, in some cases, we need quick diagnoses. So we propose an image segmentation model to predict and give the output a mask that will assist the doctor in taking this crucial decision. Deep Learning has proved their worth in many areas and outperformed man state-of-the-art models. We want to use the power of these deep learning model to solve this problem. We have used U-net [13] architecture with ResNet [17] as a backbone and achieved promising results. U-net [13] performs very well in medical image processing and semantic segmentation. Our problem falls in the semantic segmentation category.

Motivation & Objective

• To develop an automated, deep learning-based system for early and accurate detection of pneumothorax in chest X-ray images.
• To reduce reliance on expert radiologists by providing a fast, deployable model with visual interpretability through segmentation masks.
• To improve diagnostic accuracy and reduce false-negative rates common in manual interpretation of X-rays.
• To leverage deep learning for semantic segmentation to localize the exact region of pneumothorax in the lung.
• To create a model that is both computationally efficient and suitable for real-time clinical deployment.

Proposed method

• The model uses a U-Net architecture with skip connections to preserve spatial details during upsampling.
• ResNet is used as the encoder backbone, pre-trained on ImageNet to improve feature extraction and training convergence.
• The network is trained using binary cross-entropy loss and the Adam optimizer with a low learning rate (0.0001).
• Data augmentation and contrast correction are applied via OpenCV to improve robustness and generalization.
• Post-processing with thresholding is applied to clean noisy predictions and generate sharp segmentation masks.
• The model is evaluated using Dice coefficient and Intersection over Union (IoU), standard metrics for segmentation similarity.

Experimental results

Research questions

• RQ1 Can a deep learning-based semantic segmentation model accurately localize pneumothorax in chest X-ray images?
• RQ2 How does integrating ResNet as a backbone improve segmentation performance in the U-Net architecture for this task?
• RQ3 To what extent can the model reduce false-negative diagnoses compared to manual reading by radiologists?
• RQ4 Can the model produce interpretable segmentation masks that support clinical decision-making?
• RQ5 How do Dice coefficient and IoU scores correlate with visual segmentation quality in this medical imaging context?

Key findings

• The model achieved a Dice coefficient of 84.3% and IoU of 82.6%, indicating strong similarity between predicted and ground-truth masks.
• The use of pre-trained ResNet weights significantly accelerated training and improved segmentation performance.
• Visual results show that predicted masks closely align with expert-annotated ground truth across diverse patient cases.
• Post-processing with thresholding effectively reduced noise and enhanced mask sharpness, improving clinical interpretability.
• The model demonstrates strong generalization and robustness, with consistent performance across test images.
• The framework is suitable for real-time deployment due to efficient inference on GPU hardware.