Image learning and correction as a means of evaluating and fault-finding computed tomography scanner image artefacts

Abstract

The concept of image reconstruction modelling in both industry and academia has led to the investigation of different artefacts found in the medical fraternity. This concept is aimed at optimising the efficiency of fault-finding in the CT scanner to automatically identify artefacts without any human intervention. It is for such reasons that an automated artefact detection model based on machine learning and vision needs to be investigated, developed, and implemented to evaluate the feasibility and reliability of such a system in a live environment setup for service engineers and specialists alike. During the image acquisition process, images were taken from Toshiba medical equipment for image processing and machine learning purposes. The data collected contain a total of 170 images namely Dataset_1 with each subset dataset comprising 85 ring and 85 metal artefact images for data testing with 19 iterations. Additionally, a subset image dataset comprised of a total of 88 image dataset namely Dataset_2 with each subset dataset having 44 ring and 44 metal artefact images. This was conceptualised in this format for validation of the model results. The iteration determinant was the prerogative of the research based on the satisfactory output results. Training of the model was created utilising Keras to allow the model to learn by identifying the different features of the image. The first test was to observe the data accuracy in terms of artefact detection and the second test was based on the overall system accuracy based on different epochs and image dataset sizes. The first model was based on a fixed 50-epoch test. The results demonstrate that the more the input dataset, the higher the accuracy with more than 90% accuracy reading. However, to prove the robustness of the model technically and scientifically, the epoch and iteration change were necessary. Subsequently, the model data loss demonstrates that the lesser the input data within the model, the higher the chances for data loss at more than 80%. The more tests and trains were conducted the lesser the data loss at below 20% for 50 epochs tests. Model two demonstrated that for 25 epochs tests based on 170 input image datasets, the test results are 87% accurate with 49% data loss from 89% accuracy and 21% data loss observed during the training phase. Subsequently, for 50 epoch tests, the results demonstrate 91% accuracy with 15% data loss, compared to 94% accuracy and 21% loss. These results thus demonstrate the feasibility of the development of an artefact-based recognition model. Furthermore, the model based on the results was able to accurately differentiate between the metal and ring artefact without any human intervention. Transfer learning was applied using three models namely the VGG16, Resnet50 and Inception_V3 algorithms. The results demonstrate the high loss of data on the transfer learning as opposed to the Custom CNN model. The models were then evaluated based on their efficiency and accuracy threshold, and the results demonstrate that transfer learning has higher accuracy than the custom CNN model as a result of the application of the custom CNN on an already-trained model and utilisation of that result dataset to the three evaluated transfer learning algorithms. This research study`s primary focus was to investigate the feasibility of developing an intelligent model with the capability for automatic artefact detection and the ability to distinguish the investigated artefacts. Post the development and test stage of the initial model, an alternative model based on transfer learning was modelled using VGG16 algorithm. The results demonstrate that these models are both sustainable and reliable. In Addition, two more models were evaluated against the two sets of “primary algorithms” namely the custom CNN model and VGG16 algorithms. The Inception_v3 and ResNet50 algorithms were added to the set of algorithms with the primary aim of comparing the results against the custom CNN and the VGG16 model. It is therefore seen that VGG16 performs better on the data accuracy aspect, where the dataset can be recognised at a higher percentage by the VGG16 model on both the test and train models as compared to the other three tested and trained models. In retrospect, the ResNet50 algorithm performs the lowest when it comes to the losses experienced by the model against the other three models. Therefore, this study concludes that the medical fraternity must invest in the advancements of automated artefact detection models utilising this research study as a base reference model based on the results output. The results presented in this research study thus demonstrate the robustness of the model and minimal differentiation in the results outputs between two developed models of less than 2% on the 88-image and 170-image datasets.