Deep Learning-based Identification of Anatomical Landmarks

Sammanfattning

The identification of anatomical structures in Magnetic Resonance Imaging (MRI) at AMRA Medical is a time-consuming process performed manually by experts. To calculate AMRA's biomarkers, specific anatomical landmarks must first be precisely located within the skeleton. This thesis investigates the use of deep learning techniques to automate the localisation of six landmarks in whole-body MRI volumes, comparing the 3D Convolutional Neural Network (CNN) nnLandmark against two variants of the State Space Model (SSM) and CNN dual-stream hybrid nnMamba. Furthermore, the study evaluates the effects on landmark accuracy when adding auxiliary input channels that consist of different anatomical segmentations for the spine and skeleton. In addition, the study explores how consistent the models are when localising landmarks across scans of the same patient from different time points. Finally, the models impact on the volumetric accuracy of AMRA's biomarker is measured. The models were trained and evaluated on a subset of the UK Biobank data set. To improve landmark detection in SSMs when using additional auxiliary masks, an alternative dual-stream approach was implemented. The results demonstrate that the baseline nnMamba variants outperform nnLandmark, with nnMamba-stem achieving the best overall baseline performance, with a $Z$-axis Mean Radial Error (MRE) of 1.30 mm and a 12 mm Success Detection Rate (SDR) of 99.4\%. When incorporating auxiliary masks, performance depended heavily on mask configuration; the labelled spine mask improved nnLandmark accuracy, whereas the sparse spine configuration reduced performance across all models. Skeletal auxiliary input masks improve time point consistency across repeated patient scans. Finally, the dual-stream nnMamba-stem model provided the most stable calculations for the Abdominal Subcutaneous Adipose Tissue (ASAT) biomarker. In conclusion, the evaluated deep learning architectures demonstrate that automatic anatomical landmark detection in the spine and femurs is both accurate and consistent, successfully computing ASAT biomarkers within AMRA’s required threshold of 0.2L. However, edge cases and complex anatomical anomalies still occur. Therefore, a degree of human verification is still a necessary step in the clinical pipeline to guarantee high-quality predictions.