2024/02/07 | Research | Artificial Intelligence
The AI in Health and Nutrition lab conducted performance tests on a computer-aided diagnosis (CAD) system for lung fibrosis, employing multi-layer-perceptron-mixers (MLP-mixers) for segmenting lung and airway anatomy. Additionally, the system identifies idiopathic lung disease patterns in a second step. Currently, MLP-mixers segmentation demonstrates performance comparable to nnU-Net on chest CT images.
In collaboration with the Department of Radiology of the Inselspital and other Swiss and international Radiology departments, the team examined the impact of MLP-mixers on a CAD system and the differences between 2D vs. 3D segmentation on segmentation accuracy and diagnostic support. The CAD’s performance underwent validation against four independent chest radiology specialists.
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 comparable performance to human readers in most cases. For the task of ILD pattern segmentation, similar results were achieved with 3D data and 2D tomography slices.
The group anticipates that further exploration of MLP-Mixer application for medical image segmentation could enhance their performance beyond the level of state-of-the-art UNets. Consequently, they plan to incorporate additional auxiliary segmentation losses at different resolutions to gradually fine-tune the final segmentation map in future experiments.
Early access to the study
Artificial Intelligence in Health and Nutrition lab