Synopsis

Keywords: Machine Learning/Artificial Intelligence, Segmentation

Whole-Body Diffusion Weighted Imaging (WBDWI) requires automated tools that delineate malignant bone disease based on high b-value signal intensity, leading to state-of-the-art imaging biomarkers of response. As an initial step, we have developed an automated deep-learning pipeline that automatically delineates the skeleton from WBDWI. Our approach is trained on paired examples, where ground truth is defined through a set of weak labels (non-binary segmentations) derived from a computationally expensive atlas-based segmentation approach. The model showed on average a dice score, precision and recall between the manual and derived skeleton segmentations on test datasets of 0.74, 0.78, and 0.7, respectively.

Background

Whole-Body Diffusion Weighted Imaging (WBDWI) is an established technique for staging and non-invasive response assessment of bone disease in patients with Advanced Prostate Cancer (APC) and Multiple Myeloma (MM)^1,2. This technique can measure the Total Disease Volume (TDV, in millilitres) and the Apparent Diffusion Coefficient (ADC), as potential response biomarkers to systemic treatments³. However, at present this still requires tedious manual delineation of disease, requiring approximately 1-2 hours of clinician time per patient, depending on the disease volume⁴. Automated disease delineation tools are thus highly desirable and a logical first step to achieve this would be delineation of the full patient skeleton. Supervised deep learning algorithms show excellent performance for segmentation tasks, but typically consume large amounts of training data consisting of manual segmentations that are not available to WBDWI. We have developed an automated pipeline for deriving full-skeleton annotations on WBDWI using computationally expensive atlas-based segmentations to derive “weak labels”, that may in turn be used to train a much faster supervised deep learning model. Furthermore, this approach intrinsically produces estimates of skeletal segmentation uncertainty that may be used for improved model interpretability.

Methods

Patient Population
For training the deep learning model two retrospective WBDWI cohorts were used: 200 patients diagnosed with APC (Cohort A) and 46 patients diagnosed with MM (Cohort B). Every patient from the APC cohort underwent baseline and post-treatment scans acquired at one of three imaging centres (169/22/9 patients per centre). Patients with confirmed MM underwent a baseline scan at one of two imaging centres (32/14 patients per centre). Patient data from datasets A and B were split into training and validation (80:20) cohorts.

Image Acquisition
WBDWI scans were acquired using two (50/900 s/mm²) or three (50/600/900 s/mm²) b-values on a 1.5T scanner (MAGNETOM Aera/Avanto, Siemens Healthcare, Erlangen, Germany), over 4-5 stations from the skull base to mid-thigh (APC) or skull vertex to knees (MM), with each station comprising 40 slices. Echo-planar image acquisition was used (GRAPPA=2) with a double-spin echo diffusion encoding scheme applied over three orthogonal encoding directions.

Development of AI-driven model on WBDWI
Figure 1 shows the automated pipeline for annotations, training and inference of the deep learning model that delineates the skeleton from WBDWI.

• Annotations

Skeleton uncertainty maps (“weak labels”) were derived from a weighted atlas-based registration algorithm. The atlas comprised 15 single centre WBDWI scans of patients with diffuse MM (Cohort C) for which the entire skeleton was manually delineated by an expert radiologist with 3 years of experience in body MRI. Two previously described AI-based models ((i) body regions classifier and (ii) and spinal cord segmentation) were employed for optimizing the initial alignment between the atlas images and target image [5]. Computational time for the atlas-based registration was on average 4 minutes on CPU (2.4 GHz Quad-Core).

• Training

A 3D patch-based U-Net model was trained to predict skeleton uncertainty maps from the preceding phase. The networks involve a 2-channel input (i) estimated ADC map and (ii) estimated b=0 s/mm² (S0) image. Convolution layers of the U-Net model employed 16/32/64/128/256 filters, 3x3 kernel size, batch normalisation, two residual units, dropout rate of 0.3 and PReLu activation function expects for the last layer that employed a sigmoid activation function. Adam optimisation (learning rate 10­^-3) was used to minimise a binary cross-entropy loss function over 150 epochs using a batch size of 4 and patch size of 128x128x64.

Results

Figure 2 shows the skeleton uncertainty maps derived from the U-Net model for three cases in the validation dataset. Voxels classified with low and high uncertainty aligned with the patient skeleton and soft tissue background, respectively.
The inference time of the trained U-Net model was within 5 seconds with improvement in speed over the atlas-based registration of 97.92%.
To quantitatively assess the accuracy of the U-Net model, the skeleton uncertainty maps for all the images in the atlas cohort (Cohort C) were derived. A threshold of 0.45 was applied to derive a skeleton mask. The mean dice score, precision and recall between the manual and derived skeleton mask were 0.74, 0.78, and 0.7, respectively.
Skeleton masks were transferred into the ADC map to derive the ADC statistics and volume within the delineated regions. The relative error of median ADC and volume between manual and automated methods was 3.27% and 10.9%, respectively.
Figures 3 and 4 show manual and derived skeleton masks superimposed on axial and Maximum-Intensity-Projection (MIP) of b=900 s/mm² images for three cases in the atlas cohort.

Discussion

We have developed a 3D U-Net model to automatically delineate the skeleton from WBDWI. The model was trained using uncertainty maps from an automated annotations phase. Our model showed similar performance to algorithms that perform skeleton segmentation from high-resolution and high SNR images (CT, T1w and T2w) published in the literature^6-8. This approach appears useful for the towards development of bone disease segmentation.
The quantitative analysis was performed using the atlas cohort which was also used for generating the weak labels. This might be seen as a potential weakness but the original images from the atlas cohort were not used directly for training. Therefore, we argue that these results form a valid test.

Acknowledgements

This project is funded by the National Institute for Health and Care Research [i4i grant II- LA-0216-20007 part of the NIHR]. The authors would like to acknowledge Mint Medical^®