Room: Exhibit Hall | Forum 8
Purpose: The detection of lung diseases via deep learning has gained popularity due to the public release of chest radiograph (CXR) datasets. While promising results for the detection of pneumothorax (PTX) have been reported, the effects of image resolution and treatment-associated devices have not been addressed. The reduced matrix size often required for convolutional neural networks (CNNs) is 224x224, which obscures the visual signs of PTX (fine line at the edge of the lung and loss of texture outside the lung). In addition, CXRs containing PTX often contain medical devices.
Methods: 1,314 CXRs from our institutional dataset and 1,236 CXRs from the publically available ChestX-ray14 dataset were cropped to generate two apex images, one per lung, and resized to 224x224, to address image resolution concerns. A radiologist interpreted the images, noting the presence of PTX and devices. 1,948 apex images(565PTX) were used for fine-tuning a VGG19 network. To address the effect of devices, another VGG19 network was fine-tuned using device-free images. The device-free image dataset was equal in number and PTX severity distribution to the image dataset with and without devices. Both fine-tuned networks were tested on an independent test set of 1,008 images(454PTX).
Results: The CNN fine-tuned with all apex images achieved an AUC of 0.83(95% CI:0.80,0.85) and the CNN fine-tuned with device-free apex images achieved an AUC of 0.81(95% CI:0.79,0.84) in the task of distinguishing between test images containing PTX and those without. Network visualization was performed using activation heatmaps, which indicated the pixels that contributed most to the CNNâ€™s classification.
Conclusion: The CNN fine-tuned with images with and without devices had a higher AUC and sensitivity; however, the false-positive fraction was increased. Activation heatmaps were generated to investigate the causes of increased false positives. Image resolution and devices must be addressed when using CNNs to detect PTX.
Funding Support, Disclosures, and Conflict of Interest: Funding: NIH T32 EB002103, NIH UL1 TR002389, NIH QIN U01 195564, and AAPM summer undergraduate research fellowship. 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.