AI Analysis Basic Course - Third Installment: Improvement of Image Quality and Analysis Accuracy of X-ray CT Reconstruction Images by Deep Learning

Deep learning can be trained to recognize and remove noise patterns from X-ray CT reconstruction images. A denoising convolutional neural network can learn from lower-quality and higher-quality versions of CT datasets, estimating the noise present in the lower-quality data. Once trained, the model applies this knowledge to noisy CT datasets and subtracts the estimated noise.

This approach is particularly useful for CT images acquired with short scanning times. Faster scanning normally produces greater noise, making structural boundaries harder to distinguish and potentially reducing quantitative analysis accuracy. Deep-learning processing can improve contrast-to-noise ratio and make structural features easier to separate while retaining the throughput advantages of short scans.

Deep learning can substantially reduce the scanning time required to obtain data suitable for quantitative analysis. In testing with a cylindrical battery, CT data acquired in 4 minutes and processed using a trained denoising model produced an electrode overhang measurement of 0.59 mm. This was identical to the result obtained from higher-quality data acquired over 17 minutes.

For a plastic component, a 2-minute scan processed with deep learning produced a surface-area measurement close to the reference measurement obtained from data corresponding to a 40-minute acquisition. These results indicate potential scanning-time reductions of approximately fourfold for the battery and 20-fold for the plastic component while maintaining comparable analytical accuracy.

A denoising convolutional neural network can be trained using paired CT datasets representing relatively noisy and high-quality images. The lower-quality dataset serves as the coarse training data, while the higher-quality dataset provides the target the model attempts to reproduce.

The model applies convolution and batch normalization to the coarse data to estimate its noise. The estimated noise is subtracted from the coarse image, and the network is repeatedly refined so that the resulting image approaches the high-quality training data. In this work, a two-dimensional DnCNN with a 10-layer convolution and batch-normalization kernel was trained over 20 epochs. Once trained, the model can process a CT dataset in a few seconds to a few minutes.

Two approaches can generate the coarse and fine datasets needed for model training. One approach uses two scans of the same sample: a short scan provides the noisy or coarse dataset, while a longer scan provides the high-quality or fine dataset.

An alternative requires only one high-quality acquisition. The complete set of projection images is reconstructed to create the fine dataset, while a subset of those projections is used to reconstruct a lower-quality version. For the plastic-part example, 2,400 projection images produced the fine dataset, while only 120 regularly sampled projections were used for the coarse dataset. Using one acquisition for both datasets also reduces potential discrepancies caused by sample drift between separate scans.

Deep learning can improve the ability to distinguish internal battery components and make quantitative measurements from rapidly acquired CT data. One important application is measuring electrode overhang, which is relevant because winding misalignment can contribute to internal short circuits and potentially lead to fire.

For a cylindrical battery, deep-learning processing increased the contrast-to-noise ratio from 5.2 in the original 4-minute dataset to 38.2. Noise reduction also allowed the electrode shape to be extracted more accurately. The measured overhang changed from 0.64 mm before processing to 0.59 mm afterward, matching the 0.59 mm measurement obtained from a 17-minute high-quality scan.

Short CT scans can introduce noise that makes the surface of a plastic component appear artificially rough. Because surface-area calculations incorporate this apparent roughness, noise can cause the calculated area to be larger than its actual value.

For a plastic component scanned for 2 minutes, the calculated surface area was 2,956 mm² before deep-learning processing. After processing, it decreased to 2,821 mm², much closer to the 2,836 mm² value obtained from the high-quality reference dataset. The contrast-to-noise ratio simultaneously increased from 10.6 to 280. This demonstrates that denoising can improve both image quality and quantitative measurements used to evaluate characteristics such as surface roughness and molding defects.

A model does not necessarily have to be trained and applied using the same CT instrument. Deep-learning models are expected to generalize to sufficiently similar samples and datasets, creating the possibility of using very high-quality data from one system as training data for images acquired elsewhere.

One potential approach is to train a model using high-quality synchrotron CT data and then apply that model to CT data collected with laboratory or manufacturing instruments. This could potentially improve laboratory or production CT datasets toward the quality obtainable at synchrotron facilities, although the demonstrated battery and plastic examples used the same CT instrument for both training and inference.