Purpose: To train and investigate performance of a novel implementation of deep U-Net architectures in the localization-based identification of pneumothorax on a chest radiograph (CXR) dataset. Typically, pneumothorax is detected by a radiologist interpreting the CXR.
Methods: The database included frontal 240 CXRs from our institution. Using anatomical landmarks denoted by a radiologist, the left and right lungs were cropped to the superior third of each lung, yielding two apex images from each CXR (480 images). To serve as localization â€œtruth,â€? a radiologist drew contour lines along the upper lung margin in cases of pneumothorax. Binary truth masks, identical in size to the cropped images, were made by widening these contour lines to four pixels in width. The 480 image dataset was randomly split by case into three groups: 65% (312) used for training the network, 20% (96) for validation, and 15% (72) for testing the performance of the network. A U-Net neural network was trained from scratch with step-wise validation. Random jitter augmentation enhanced the training, which involved 33 epochs using an Adam optimizer, an adaptive learning rate, 0.05 dropout, and per-layer batch normalization. Subsequently, a scanning region-of-interest (ROI) method was used to localize the U-Net indicated pneumothorax.
Results: During evaluation, the total pixel value sum of each scanning ROI was calculated, assigning a per-image image score as the largest ROI value in that image. ROC analysis of the largest ROI pixel sums yielded an area under the curve of 0.84 [95% confidence interval:(0.73,0.92)].
Conclusion: Our novel implementation of a U-Net architecture combined with a scanning ROI for localization demonstrated good performance at localization-based identification of pneumothorax in frontal CXRs. Ongoing research will address the localization performance and generalizability of the network, as well as the impact of medical devices on classification and test on an independent dataset.
Funding Support, Disclosures, and Conflict of Interest: Funding: AAPM summer undergraduate research fellowship (TR), NIH T32 EB002103, NIH UL1 TR002389, NIH QIN U01 195564. MLG: stockholder in R2 Technology/Hologic, a cofounder and equity holder in Quantitative Insights, and a shareholder in QView. MLG receives royalties from Hologic, GE Medical Systems, MEDIAN Technologies, Riverain Medical, Mitsubishi, and Toshiba.
Not Applicable / None Entered.