Medical image segmentation is a crucial element of computer-aided diagnosis (CAD) systems. Segmentation maps are used to calculate imaging features, such as quantitative disease distribution and radiomic features. Since their introduction in 2015, UNets have become the state-of-the-art segmentation tools. However, since that time, many new methods for image processing have been introduced, such as vision transformers and multi-layer-perceptron-mixers (MLP-Mixers). Alongside baseline UNets, we have now investigated the application of such MLP-Mixers for medical image segmentation, as part of a CAD system for the diagnosis of interstitial lung diseases (ILDs). Furthermore, we have investigated the effect of 2D and 3D data representations on segmentation and the final CAD results. We have evaluated the performance of the baseline segmentation methods and the MLP-Mixer primary on the overall diagnostic performance of the CAD system - as well as on the accuracy of segmentation as an intermediate step. In addition to network and data representation variations, we have investigated two different techniques for selecting features, an agnostic method and an alternative approach which selects features tailored to a specific segmentation map and diagnosis task. Finally, the CAD’s performance was compared with that of four independent specialists in chest radiology. Among the 105 test cases, the diagnostic accuracy was 77.2±1.6% for the AI-approaches and 79.0±6.9% for the radiologists, indicating that the proposed systems perform comparably well to human readers in most of the cases. For the task of ILD pattern segmentation, similar results were obtained with 3D data and 2D tomography slices.
