Bone Suppression on Chest Radiographs for Pulmonary Nodule Detection: Comparison between a Generative Adversarial Network and Dual-Energy Subtraction

Abstract

Objective: To compare the effects of bone suppression imaging using deep learning (BSp-DL) based on a generative adversarial network (GAN) and bone subtraction imaging using a dual energy technique (BSt-DE) on radiologists' performance for pulmonary nodule detection on chest radiographs (CXRs).

Materials and methods: A total of 111 adults, including 49 patients with 83 pulmonary nodules, who underwent both CXR using the dual energy technique and chest CT, were enrolled. Using CT as a reference, two independent radiologists evaluated CXR images for the presence or absence of pulmonary nodules in three reading sessions (standard CXR, BSt-DE CXR, and BSp-DL CXR). Person-wise and nodule-wise performances were assessed using receiver-operating characteristic (ROC) and alternative free-response ROC (AFROC) curve analyses, respectively. Subgroup analyses based on nodule size, location, and the presence of overlapping bones were performed.

Results: BSt-DE with an area under the AFROC curve (AUAFROC) of 0.996 and 0.976 for readers 1 and 2, respectively, and BSp-DL with AUAFROC of 0.981 and 0.958, respectively, showed better nodule-wise performance than standard CXR (AUAFROC of 0.907 and 0.808, respectively; p ≤ 0.005). In the person-wise analysis, BSp-DL with an area under the ROC curve (AUROC) of 0.984 and 0.931 for readers 1 and 2, respectively, showed better performance than standard CXR (AUROC of 0.915 and 0.798, respectively; p ≤ 0.011) and comparable performance to BSt-DE (AUROC of 0.988 and 0.974; p ≥ 0.064). BSt-DE and BSp-DL were superior to standard CXR for detecting nodules overlapping with bones (p < 0.017) or in the upper/middle lung zone (p < 0.017). BSt-DE was superior (p < 0.017) to BSp-DL in detecting peripheral and sub-centimeter nodules.

Conclusion: BSp-DL (GAN-based bone suppression) showed comparable performance to BSt-DE and can improve radiologists' performance in detecting pulmonary nodules on CXRs. Nevertheless, for better delineation of small and peripheral nodules, further technical improvements are required.

Keywords: Bone suppression imaging; Chest radiography; Deep learning; Generative adversarial network; Pulmonary nodules.

Fig. 1. Flow chart depicting the selection of the study population.

CXR = chest radiograph

Fig. 2. The architecture of bone suppression algorithm based on wavelet transform and generative adversarial networks.

A. The architecture of generator that receives the original image and produces bone suppressed images. This system uses the frequency information obtained from Haar wavelet transformation which pre-defines features that the network should learn, allowing the network to converge more quickly and efficiently. The generator takes four channels of 512 × 512 source (original) images obtained through Haar 2D wavelet decomposition and tries to produce output (bone suppressed) image that can fool the discriminator by avoiding image blurring. The output image is finally reconstructed to 1024 × 1024 by Haar 2D wavelet reconstruction. Values below each conv block are image compression ratios with the number of channels. B. The architecture of the discriminator. The discriminator distinguishes whether the input is a fake image that comes from the generator, or a real one from the training set. In the training process, the two networks compete. The discriminator plays a critical role in preventing the generator from producing a blurred image and also considers the distribution of image batches s as the generator does. conv = convolution, fc = fully connected, 2D = two-dimensional

Fig. 3. A 27-year-old female who presented with cough.

A. A nodule in the right pericardiac area (arrow) is obscured by rib shadow on the CXR. B, C. BSp-DL (B) and BSt-DE (C) reveal the nodule (arrows) clearly. In the evaluation of image quality, readers scored visibility of pulmonary vessels in the lung fields higher in BSt-DE while visibility of pulmonary vessels in cardiac and diaphragmatic areas was higher in BSp-DL (circles). D. A coronal image of chest CT in lung window setting showing a 16 mm nodule (arrow) in the right lower lobe. BSp-DL = bone suppression imaging using deep learning, BSt-DE = bone subtraction imaging using dual energy technique, CXR = chest radiograph

Fig. 4. The AFROC curves showing nodule-wise localization performance of radiologists.

Area under the AFROC curve is improved significantly (p < 0.017) using BSp-DL or using BSt-DE than that using CXR alone. AFROC = alternative free-response receiver operating characteristic, BSp-DL= bone suppression imaging using deep learning, BSt-DE = bone subtraction imaging using dual energy technique, CXR = chest radiograph, R1 = reader 1, R2 = reader 2

Fig. 5. An 85-year-old male who underwent CXR as a part of routine follow-up after gastrectomy for stomach cancer.

A. A tiny nodule overlapping with anterior arc of right 3rd rib (arrow) is suspicious in the right upper lobe on CXR. B, C. BSp-DL (B) and BSt-DE (C) confirm persistence of the nodule (arrows) after eliminating ribs. D. A coronal image of chest CT in lung window setting showing the presence of a 7 mm nodule (arrow) in the right upper lobe. BSp-DL= bone suppression imaging using deep learning, BSt-DE = bone subtraction imaging using dual energy technique, CXR = chest radiograph

Fig. 6. A 61-year-old male who presented with cough and sputum.

A. Readers missed a subpleural nodule in the left lower lobe near the costophrenic angle on chest radiograph. B. BSp-DL cannot define the nodule clearly. C. BSt-DE delineates the nodule (arrow) in the left lower lobe. D. A coronal image of chest CT in lung window setting showing the presence of a 9 mm nodule (arrow) in the left lower lobe. BSp-DL= bone suppression imaging using deep learning, BSt-DE = bone subtraction imaging using dual energy technique

Fig. 7. A 77-year-old male with prostate cancer.

A. CXR shows a nodule opacity in the right lower lung zone due to an old fracture of the right 5th rib. B. BSp-DL showing no parenchymal lesion except the same opacity seen in CXR. C. BSt-DE showing additional nodular opacity in the left upper lobe (arrow). D. A coronal image of chest CT in lung window setting showing no parenchymal nodule in left upper lobe, but hypertrophied costochondral junction of the left 1st rib (arrow). BSp-DL= bone suppression imaging using deep learning, BSt-DE = bone subtraction imaging using dual energy technique, CXR = chest radiograph