Spatial resolution enhancement using deep learning improves chest disease diagnosis based on thick slice CT

Abstract

CT is crucial for diagnosing chest diseases, with image quality affected by spatial resolution. Thick-slice CT remains prevalent in practice due to cost considerations, yet its coarse spatial resolution may hinder accurate diagnoses. Our multicenter study develops a deep learning synthetic model with Convolutional-Transformer hybrid encoder-decoder architecture for generating thin-slice CT from thick-slice CT on a single center (1576 participants) and access the synthetic CT on three cross-regional centers (1228 participants). The qualitative image quality of synthetic and real thin-slice CT is comparable (p = 0.16). Four radiologists' accuracy in diagnosing community-acquired pneumonia using synthetic thin-slice CT surpasses thick-slice CT (p < 0.05), and matches real thin-slice CT (p > 0.99). For lung nodule detection, sensitivity with thin-slice CT outperforms thick-slice CT (p < 0.001) and comparable to real thin-slice CT (p > 0.05). These findings indicate the potential of our model to generate high-quality synthetic thin-slice CT as a practical alternative when real thin-slice CT is preferred but unavailable.

Fig. 1. Overview of deep learning synthetic model: a convolutional-transformer hybrid encoder-decoder architecture synthesizes thin-slice CT from thick-slice CT by recovering masked regions from visible regions.

a The Encoder maps the input L slices from the original thick-slice CT (visible regions) to a latent representation. Masked regions are introduced via the Mask Token Add Module and combined with the latent representation. The Decoder then recovers the masked regions from the latent representation, producing an output size of 5 × (L–1) + 1 through the final Linear Projection. b The CTH Block comprises four successive STLs and a Conv. The 3D CTH Block consists of 3D STL and 3D Conv, while the 2D CTH Block consists of 2D STL and 2D Conv. c The T-CTH Block has two parallel branches that perform feature extraction from the coronal and sagittal views, respectively. The permutation operation P is used to transform the input view to coronal or sagittal views, or vice versa. d Details of two successive 2D or 3D STLs. CTH Block indicates convolutional-transformer hybrid block; T-CTH Block, through-plane convolutional-transformer hybrid block, Conv convolutional, P permutation operation, W-MSA window multi-head self-attention, SW-MSA shift window multi-head self-attention, MLP multi-layer perceptron.

Fig. 2. Different CT images for 24-year-old man from dataset-development.

a Axial view displayed as the lung window. b Coronal view displayed as the lung window. c Sagittal view displayed as the bone window. BIS indicates bicubic interpolation synthetic; DLS deep learning synthetic.

Fig. 3. Different CT images for 26-year-old woman from dataset-USA.

a Axial view displayed as the lung window. b Coronal view displayed as the lung window. c Sagittal view displayed as the bone window. BIS indicates bicubic interpolation synthetic, DLS deep learning synthetic, CT computed tomography, USA United States of America.

Fig. 4. Stacked bar graphs display the distribution of quality scores.

Eight radiologists independently rated Real, BIS, and DLS 1-mm CT using a five-point Likert scale (1 = unacceptable, 2 = poor, 3 = acceptable, 4 = good, 5 = excellent). In the Likert scale, scores of ‘unacceptable’ and ‘poor’ are defined as nondiagnostic (displayed in varying shades of red); scores of ‘acceptable’, ‘good’ and ‘excellent’ are defined as diagnostic (displayed in varying shades of green). BIS indicates bicubic interpolation synthetic, DLS deep learning synthetic, CN China, US United States.

Fig. 5. Diagnostic Evaluation of Lung Nodule Detection.

Sensitivity of each reader using various types of CT images, shown for a all nodules (N = 200), b solid nodules (N = 132), c calcific nodules (N = 52), and d subsolid nodules (N = 16). DLS indicates deep learning synthetic, R reader, CT computed tomography.

Fig. 6. Evaluation of AI-assisted diagnostic product.

a ROC curves of AI-assisted CAP diagnosis with different CT images. b Sensitivity of AI-assisted lung nodule detection with different CT images. AI indicates artificial intelligence, ROC receiver operating characteristic, CAP community-acquired pneumonia, DLS deep learning synthetic, BIS bicubic interpolation synthetic, CT computed tomography.