Fully automated longitudinal segmentation of new or enlarging Multiple Scleroses (MS) lesions using 3D convolution neural networks

Zielsetzung The quantification of new and enlarging Multiple Scleroses (MS) lesions from follow-up MRI scans is an important surrogate of clinical disease activity. Manual assessment is time consuming, inter-rater (IR) variability is high, and only few fully automated methods are available so far.

Material und Methoden A 3D convolution neural network (CNN) with encoder-decoder (U-Net-like) architecture was employed. Input data consisted of two fluid attenuated inversion recovery (FLAIR) images (baseline (BL) and follow-up (FU)). Each image was entered into the encoder independently and the feature maps were concatenated and fed into the decoder. The output was a 3D mask indicating new (in FU) or enlarged (compared to BL) lesions. The encoder was pre-trained on cross-sectional MS patient data using 1,809 2D and 3D FLAIR images acquired on 156 different scanners. MRI data originated from clinical routine and was sent to jung diagnostics GmbH for image analysis. For the evaluation and the training of the decoder (with leave-9-out cross validation of 10 nets), 89 RRMS patient’s data was acquired at the University Hospital of Zurich. For each patient a BL and FU 3D FLAIR image (3T Philips Ingenia) were used (mean follow-up 2.2 years). New or enlarging lesions were annotated on the FU scans by two raters.

Ergebnisse Rater 1 and 2 identified on average 1.1 new or enlarging lesions (interquartile range (IQR) [0.0,1.0]). For evaluation, mean and IQR of lesion-wise sensitivity (SEN) and false positive count (FP) were determined. We obtained the following results between raters (IR) and between raters and CNN (R-CNN): IR: SEN: 0.60 [0 1] FP: 0.753 [0 1]; R-CNN: SEN: 0.58 [0.33 1] FP: 0.625 [0 1].

Schlußfolgerungen The low inter-rater performance signifies the complexity and uncertainty of identifying new and enlarging lesions. An automated CNN-based approach can quickly (<1 min) provide an independent and deterministic assessment of lesions from BL and FU scans to support diagnosis and potentially mitigate inter-rater variability.