[Back]

R. Licandro, J. Hofmanninger, M. Perkonigg, S. Röhrich, M. Weber, M. Wennmann, L. Kintzele, M. Piraud, B. Menze, G. Langs:
"Evolution Risk Prediction of Bone Lesions in Multiple Myeloma";
Poster: European Congress of Radiology, Wien; 2020-07-15 - 2020-07-19.

Purpose
The earliest possible detection of bone lesions is key to facilitate timely treatment decisions in patients with Multiple Myeloma (MM). The clinical relevance lies in providing a machine-learning based score to assess the risk of localized bone regions to evolve into diffuse or osteolytic lesions based on pre-stage infiltration patterns in whole body Magnetic Resonance Imaging (wbMRI).

Materials and Methods
63 patients received multiple T1 weighted wbMRI scans with the average time of 13 months between the scans. Overall, 170 locations evolved to either diffuse or osteolytic lesions. Here, we propose a methodology to predict future bone lesion growth and emergence from T1 weighted wbMRI images resulting in a full body risk score map. We propose an asymmetric cascade architecture of U-Nets consisting of an MR image-based bone segmentation net and a patch-based lesion prediction net. The algorithm identifies early signatures of emerging lesions and visualizes high risk locations accordingly.

Results
The proposed approach is evaluated for two body parts ((1) thorax, (2) legs) and lesion types. The bone segmentation net detects bones with a mean Area Under the Curve (AUC) of 0.76423 (1) and 0.8023 (2). The lesion prediction net predicts emerging lesions (which are reported in the future but not in the observed scan), with a mean AUC of 0.6083 (1) and 0.5304 (2) and changing lesions (which are annotated continuously) with a mean AUC of 0.5855 and 0.6840.

Conclusion
We propose a risk predictor for lesions to emerge or to progress and map high-risk regions accordingly. We first segment bones and then predict lesions within bones using asymmetric cascaded U-Nets. This is the first approach that predicts lesions on full volumetric wbMRI data, showing feasible results, but false positives occur in areas of anomaly, that do not progress to lesions. Our findings indicate hidden imaging markers beyond lesion size, currently used for categorization and risk stratification in MM.

Purpose
The earliest possible detection of bone lesions is key to facilitate timely treatment decisions in patients with Multiple Myeloma (MM). The clinical relevance lies in providing a machine-learning based score to assess the risk of localized bone regions to evolve into diffuse or osteolytic lesions based on pre-stage infiltration patterns in whole body Magnetic Resonance Imaging (wbMRI).

Materials and Methods
63 patients received multiple T1 weighted wbMRI scans with the average time of 13 months between the scans. Overall, 170 locations evolved to either diffuse or osteolytic lesions. Here, we propose a methodology to predict future bone lesion growth and emergence from T1 weighted wbMRI images resulting in a full body risk score map. We propose an asymmetric cascade architecture of U-Nets consisting of an MR image-based bone segmentation net and a patch-based lesion prediction net. The algorithm identifies early signatures of emerging lesions and visualizes high risk locations accordingly.

Results
The proposed approach is evaluated for two body parts ((1) thorax, (2) legs) and lesion types. The bone segmentation net detects bones with a mean Area Under the Curve (AUC) of 0.76423 (1) and 0.8023 (2). The lesion prediction net predicts emerging lesions (which are reported in the future but not in the observed scan), with a mean AUC of 0.6083 (1) and 0.5304 (2) and changing lesions (which are annotated continuously) with a mean AUC of 0.5855 and 0.6840.

Conclusion
We propose a risk predictor for lesions to emerge or to progress and map high-risk regions accordingly. We first segment bones and then predict lesions within bones using asymmetric cascaded U-Nets. This is the first approach that predicts lesions on full volumetric wbMRI data, showing feasible results, but false positives occur in areas of anomaly, that do not progress to lesions. Our findings indicate hidden imaging markers beyond lesion size, currently used for categorization and risk stratification in MM.

Muskoskelatal, Computer-Aided Diagnoses, Cancer, Artificial Intelligence

http://dx.doi.org/10.1186/s13244-020-00851-0