Artificial Intelligence in Medical Image Processing
Since the breakthrough of AlexNet on the ImageNet Classification Challenge in 2012, the computer vision community developed a wide variety of architectures based on convolutional neural networks (CNN). The availability of high quality image data as well as the ever-increasing processing power of CPUs and the parallelization performance of GPUs fuel their success.
Modern medical diagnosis and treatment produces a vast amount of data: From small dermatology images, over the live video feed of a surgery or the intra-operative ultrasound to the volumetric images of a magnetic resonance imaging (MRI) or computed tomography (CT) scan. Automatic interpretation of these images allows for a speed up and more reliable diagnosis while lowering costs at the same time. Related tasks in which artificial intelligence systems excel are the segmentation and classification of images as well as regression. Research currently conducted at the Chair of Medical Engineering includes convolutional neural network architectures for ultrasound segmentation as well as context sensitive instrument identification and pose estimation.
Figure 1 Automatic segmentation of the distal femur. Ultrasound image (1st and 3rd image) compared to segmentation (2nd and 4th image) returned by a trained ‘U-Net’, a common Convolutional Neural Network architecture.
Publications
- P. Brößner, B. Hohlmann & K. Radermacher: Transformer vs. CNN – A Comparison on Knee Segmentation in Ultrasound Images. In: F. Rodriguez Y Baena, J.W. Giles & E. Stindel (ed.): Proceedings of The 20th Annual Meeting of the International Society for Computer Assisted Orthopaedic Surgery, 5, 2022, pp. 31-36 [DOI: 10.29007/cqcv]
- B. Hohlmann, P. Brößner & K. Radermacher: CNN based 2D vs. 3D Segmentation of Bone in Ultrasound Images. In: F. Rodriguez Y Baena, J.W. Giles & E. Stindel (ed.): Proceedings of The 20th Annual Meeting of the International Society for Computer Assisted Orthopaedic Surgery, 5, 2022, pp. 116-120 [DOI: 10.29007/qh4x]
- P. Pandey, B. Hohlmann, P. Brößner, I. Hacihaliloglu, K. Barr, T. Ungi, O. Zettinig, R. Prevost, G. Dardenne, Z. Fanti, W. Wein, E. Stindel, F. Arambula Cosio, P. Guy, G. Fichtinger, K. Radermacher & A. Hodgson: Standardized Evaluation of Current Ultrasound Bone Segmentation Algorithms on Multiple Datasets. In: F. Rodriguez Y Baena, J.W. Giles & E. Stindel (ed.): Proceedings of The 20th Annual Meeting of the International Society for Computer Assisted Orthopaedic Surgery, 5, 2022, pp. 148-153 [DOI: 10.29007/q51n]
- P. Broessner, B. Hohlmann & K. Radermacher: Ultrasound-based Navigation of Scaphoid Fracture Surgery. In: C. Palm, T.M. Deserno, H. Handels, A. Maier, K. Maier-Hein & T. Tolxdorff (ed.): Bildverarbeitung für die Medizin 2021, Springer, 2021, pp. 28-33 [DOI: 10.1007/978-3-658-33198-6_8]
- B. Hohlmann, J. Glanz & K. Radermacher: Segmentation of the distal femur in ultrasound images. Current Directions in Biomedical Engineering, 6(1), 2020, pp. 1-5 [DOI: 10.1515/cdbme-2020-0034]
- B. Hohlmann & K. Radermacher: Augmented Active Shape Model Search – towards 3D Ultrasound-based Bone Surface Reconstruction. In: F. Rodriguez Y Baena & F. Tatti (ed.): CAOS 2020. The 20th Annual Meeting of the International Society for Computer Assisted Orthopaedic Surgery, 4, 2020, pp. 117-121 [DOI: 10.29007/3px6]
- B. Hohlmann & K. Radermacher: The interleaved partial active shape model (IPASM) search algorithm - towards 3D ultrasound-based bone surface reconstruction. In: P. Meere & F. Rodriguez Y Baena (ed.): CAOS 2019. The 19th Annual Meeting of the International Society for Computer Assisted Orthopaedic Surgery, 3, 2019, pp. 177-180 [DOI: 10.29007/rbgl]