Synopsis

Keywords: Machine Learning/Artificial Intelligence, Cancer

Quantitative magnetic resonance imaging (qMRI) can provide additional information for diagnosis and response assessment, but adoption of multi-parametric qMRI techniques has been hindered by long acquisition times and labor-intensive processing steps. Magnetic resonance fingerprinting (MRF) provides quantitative maps in a single acquisition but MRF deployment in clinical studies still requires manual delineation of volumes of interest. A spatial-adaptive deep learning framework was developed to segment cervical cancer on MRI and quantify T₁ relaxation times of the tumor pre- and post-radiotherapy treatment. Our results suggest that automated segmentation models may be promising tools for quantitative tumor evaluation and treatment response assessment.

Background

Quantitative magnetic resonance imaging (qMRI) denotes mapping of underlying quantitative tissue properties (e.g. T₁, T₂ relaxation times)¹. Multi-parametric qMRI methods suffer from long acquisition times, which hinder integration into routine clinical care. Recently, the introduction of magnetic resonance fingerprinting (MRF), for simultaneous estimation of these tissue properties, has shifted the paradigm in quantitative imaging². While MRF promises to bring qMRI to cancer tumor analysis, it relies on manual delineation of regions of interest that increases the clinician workload. In this study, we propose a novel spatial-adaptive deep learning (DL)-based framework to automatically delineate cervical tumors on T₂-weighted and quantify changes in T₁ relaxation times between MRI pre- and post-radiotherapy (RT) measurements. Moreover, this pipeline offers an approach to bridge the semantic knowledge between two-dimensional (2D) diagnostic and three-dimensional (3D) RT planning MRI scans for improved segmentation outcome.

Purpose

Develop and evaluate a novel and end-to-end deep learning framework for automatic segmentation and quantitative analysis of cervical cancer tumors pre- and post-treatment using semantic knowledge transfer and MRF respectively.

Methods

Patient Population and Imaging Protocols
Institutional review board approval was obtained for this study. Three cohorts were identified for tumor segmentation pre-training, transfer learning and evaluation. Cohort 1: 125 patients with cervical cancer undergoing 2D diagnostic abdomen-pelvis T₂-weighted MRI (T₂W) in axial and sagittal planes (MAGNETOM Trio, Siemens Healthcare, Erlangen, Germany). Cohort 2: 21 patients with 3D T₂W MRI acquired for RT planning (MAGNETOM Aera, Siemens Healthcare, Erlangen, Germany). Cohort 3: 18 patients undergoing 3D T₂W MRI for RT planning pre-treatment, from whom 13 patients also had 2D T₂W MRI for post-treatment follow-up. The MRF measurements were acquired during both patient visits (9 axial slices, positioned at the center of cervical tumors) using a prototype Fast Imaging with Steady State Precession (FISP) sequence³ with Gadgetron reconstruction⁴ that returns maps of T₁ and T₂ relaxation times, and proton density M₀. In this cohort, 16 patients had a second baseline MRF acquisition for assessment of repeatability (Figure 1).

Framework Architecture and Training
All images in cohort 1 were normalized (mean: 0, variance: 1) and resampled to an isotropic global space to match the resolution of images in cohort 2 (0.8×0.8×0.8 mm³). The axial and sagittal planes embed global semantic information of diagnostic images to transfer to planning MR images. A modified residual U-Net⁵ with 4 down- and up-sampling blocks and 2 residual units was developed to perform 3D patch-based (128×128×128) tumor segmentation (train: 75, validation: 25, test: 25 patients). The trained weights were used for segmentation fine-tuning of cohort 2 MR images (train: 13, validation: 4, test: 4 patients). The quantitative segmentation metrics were investigated after four-fold cross validation for target-only training and pre-training segmentation adaptation. The training for post-treatment tumor segmentation was implemented on sagittal MRI images of size (256×256×16) from cohort 1 to expand the field of view and reduce pre-processing workload. The segmentation volumes were generated by sliding window patch fusion and intensity averaging (75% overlap).

Magnetic Resonance Fingerprinting (MRF) T₁ Measurements
The MRF sequence was validated on the ISMRM/NIST MRI system phantom (CaliberMRI, Inc, Boulder, CO) by comparison with reference T₁ values from the phantom manufacturer and comparison with T₁ relaxation times estimated using inversion recovery turbo spin echo (IR-TSE) (10 inversion times) non-linear curve fitting⁶. For these comparisons, the median T₁ values were extracted from the regions of interest (ROIs) delineated for each sphere. The T₁ measurements of cervical tumors in cohort 3 were based on 16 patients who had two measurements during their first visit. The tumor contours were generated by the DL framework, and manually revised to account for internal tissue motions between scans. The schematic of the framework is shown in Figure 2.

Results

The segmentation results suggest that large-scale transfer learning from 2D diagnostic to 3D planning MRI improved the segmentation outcome compared with training without domain knowledge transfer (Figure 3). The MRF repeatability measurements for the ISMRM/NIST phantom and cohort 3 segmentations are shown in Figure 4. From 13 patients that underwent both pre- and post-MRF scanning, the cervical tumors were detected for 6 patients by the framework. From preliminary analysis, T₁ relaxations in 5/6 patients were reduced after treatment. However, two-tail paired T-test statistics revealed no significant changes in these values (p>0.05) (Figure 5).

Discussion and Conclusion

The rise of DL and rapid qMRI techniques offer the opportunity for automated quantitative measurements in clinical MRI. Quantitative properties from the MRF can be promising biomarkers in characterising RT treatment response in cancer⁷. We developed an end-to-end spatial-adaptive framework whereby cervical cancer can be delineated automatically and propagated to MRF scans for quantitative analysis. The key limitations are internal motion during the examination, leading to differences in tumor/organ position between the anatomical and qMRI data, and difficulty in segmenting some tumors. Another drawback is that the MRF acquisition consisted of nine 5-mm slices, due to constraints on acquisition time, and therefore did not cover the whole tumor in some cases. Future work will include qualitative radiologist assessments and automatic quantification of apparent diffusion coefficients (ADC) of cervical cancers. Additionally, multi-parametric MRI input for segmentation training will be investigated.

Acknowledgements

This study represents independent research funded by the National Institute for Health and Care Research (NIHR) Biomedical Research Centre and the Clinical Research Facility in Imaging at The Royal Marsden NHS Foundation Trust and The Institute of Cancer Research, London, United Kingdom. The views expressed are those of the author(s) and not necessarily those of the NIHR or the Department of Health and Social Care. Gigin Lin received research funding from the Ministry of Science and Technology Taiwan (MOST 110-2628-B-182A-018). We are grateful to Vikas Gulani and Yun Jiang at the University of Michigan for supplying the MRF sequence.