Fast detection of cracks in ultrasonically welded parts by inline X-ray inspection

Individual inspection of medical devices is crucial to guarantee high quality products but it is also very demanding. Therefore, in this work, a next generation inline approach is proposed for inspection of ultrasonically welded parts with X-rays in a production line. The focus of the inspection is in the detection of cracks at the welded interface of these parts. The approach consists of a two-step procedure in which first, images are reconstructed from a small number of X-ray radiographs and secondly, crack detection is applied. Results show that implementation of the inline inspection during the production process is feasible in terms of time and quality constraints. Inline inspection, neural networks, filtered backprojection, classification


Introduction
In healthcare, the correct functioning of all medical devices is crucial for patient comfort and even to save lives.However, in many cases it is currently only possible to inspect some samples of a batch to determine the overall batch quality.In this abstract, we propose a next generation inline inspection approach for medical devices, with a focus on crack detection in ultrasonically welded parts [1].With the inline approach, every device is inspected individually with X-rays and is classified as 'good' or 'cracked' based on the tomographic reconstructions.The approach reduces the number of cracked devices passing the inspection as well as the number of correct devices failing it.Important parameters for the approach to be implemented in the production chain are the speed of the data acquisition, reconstruction and classification as well as their quality.In this abstract, first the proposed method is described, after which some preliminary results are shown.

Method
In the inline inspection approach under concern, all samples to be inspected are placed behind each other on a rotating holder on a conveyor belt.At regular time points, the conveyor belt is halted and the parts in the field of view of the source and the detector start to rotate during which projections from several projection directions are taken.After the data acquisition, the conveyor belt starts moving again, new parts enter and the process repeats itself.Subsequently, the crack detection algorithm is performed on the acquired data.The algorithm exists of two steps.In the first step, 2D images of slices through the parts are reconstructed with an adapted version of the NN-hFBP algorithm for the specific geometry [2].Secondly, image processing tools are applied on the reconstructed images to classify the parts as 'good' or 'cracked'.In the image processing step, the images are thresholded to detect pixels that possibly indicate a crack.By taking into account small regions around the selected pixels and studying the presence of cracks in successive images, the classification is performed.To make the inline inspection feasible, both steps are performed very fast.

Experiments and Results
To validate the proposed inline inspection approach, 51 ultrasonically welded parts of which 30 contained cracks at the welded interface were scanned by XRE nv with 1500 X-ray projections over a range of 360 degrees.In the acquisition geometry, the device was placed on the line between the source and detector so that only one device was scanned at the same time.However, the method is easily extendible to multiple devices.FBP reconstructions of 60 slices around the welded interface were made with the ASTRA toolbox [3] and used as ground truth images for the NN-hFBP reconstruction algorithm.Because the cone-angle was only 0.86 degrees, the fan-beam reconstruction algorithm could be applied.To create the 2D reconstruction images, for different numbers of projections ranging between 10 and 150, several networks were trained.For each network, the parts with cracks were subdivided into 20 parts for training, five for validation and five for testing.Of these parts, 1.000.000pixels of 600 images were used for training and 10.000 pixels of 100 images for validation.A smart selection of the pixels was proposed by using crack maps that indicate the regions of the cracks.To evaluate the quality of the NN-hFBP reconstructions, the Root Mean Squared Error (RMSE) between the reconstructed images and the reconstructions with 1500 projections was calculated for an increasing number of projections in Figure 1.The figure shows the RMSE normalized by the number of pixels both in the region surrounding the cracks and in the whole adapter region for NN-hFBP reconstructions where the network was trained on the crack maps (CM) and on the whole image (Im) as well as for FBP reconstructions.The quality of the reconstructions made with the NN-hFBP algorithm trained on the crack maps is slightly better than the one trained on the whole image and clearly outperforms the quality of the FBP reconstructions.Table 1 shows the NN-hFBP reconstruction time and figure 2 shows NN-hFBP reconstruction images made with different number of projections when the network is trained with the crack maps.In this work, we subsequently performed the classification on FBP and NN-hFBP reconstructions made with 50 projections.The results are compared in Table 2 to results obtained when applying the classification algorithm on the reconstructions made with 1500 projections.The precision and accuracy are clearly better for the NN-hFBP reconstructions compared to the FBP reconstruction.The classification time is also shown in Table 2.

Conclusion
The proposed inline inspection approach exists of a two-step procedure where first NN-hFBP reconstructions of the parts are made after inline scanning and secondly an image processing algorithm is applied for classification.The performance of the approach for medical devices is evaluated and its feasibility in practice is demonstrated by the short reconstruction and classification time.It is shown that when only 50 projections are acquired, a precision of 0.85 and accuracy of 0.80 can be obtained.The major advantage of the new approach is the fast individual inspection that can be fit into the production chain.

Table 1 :
NN-hFBP Reconstruction time in function of the number of projections.

Figure 1 :
Figure 1: normalized RMSE of the reconstructed images in function of the number of projections.

Table 2 :
Precision, accuracy, recall, specificity and classification time of the classification algorithm after application on a high-quality dataset made with 1500 projections and on the two reconstructed dataset made with FBP and NN-hFBP with 50 projections.