AI Networks Could Predict Aggressive Brain Tumour Outcomes

Liam Davenport

November 04, 2020

A deep learning neural network that rapidly calculates the cross-sectional area (CSA) of the temporalis muscle on standard magnetic resonance (MR) scans could be used to predict survival outcomes in patients with the aggressive brain malignancy glioblastoma, shows UK research.

“We realised that sarcopenia could be identified by quantifying muscle in cross-sectional imaging that cancer patients routinely undergo,” said lead researcher Dr Ella Mi, a clinical research fellow at Imperial College London, in a news release.

“This would allow for opportunistic screening of sarcopenia as part of cancer care without additional scanning time, radiation dose or cost.”

The research was presented at the National Cancer Research Institute (NCRI) Virtual Showcase 2020 on November 2.

“Doctors who treat cancer patients know that a person’s physical condition and frailty is linked to their ability to tolerate treatment,” David Harrison, professor of pathology at the University of St Andrews and chair of the NCRI Cellular Molecular Pathology Initiative, commented.

“It also affects how they respond and their long-term survival.”

He added: “However, it has been very difficult to develop objective measurements of patients’ performance status.”

Professor Harrison, who was not involved in the study, said that, at this stage, the current findings “show that there is an association between the temporalis muscle size and a patient’s frailty and how their disease progresses”.

However, he cautioned that the results “do not show that the muscle area actually causes the change in patients’ outcomes” and called for further study “in a larger group of patients to take account of factors such as changes to the muscle caused by other factors, including surgery or radiotherapy”.

Prostate Cancer

In a second study presented at the meeting, researchers led by David Gillespie, from Manchester Metropolitan University, applied deep learning networks to three publicly available datasets of MR imaging in 132 prostate cancer patients.

They found that, in their proof of concept study, that assessment of the gross tumour volume (GTV) with non-proprietary technology was not only feasible but the results were comparable to those achieved by human assessment.

First Study 

Presenting the first study, Dr Ella Mi said that glioblastoma is associated with a 5-year survival of less than 5%.

While performance status is known to be a prognostic factor for cancer patients, it is a subjective assessment. In contrast, sarcopenia, which is also a poor prognostic factor, can be quantified on the kind of cross-sectional imaging that is routinely performed on cancer patients.

Mi pointed out, however, that the manual measurement or segmentation of muscle is time intensive, susceptible to inter-rater inconsistency and unrealistic for the assessment of large datasets.

The team therefore developed a deep learning system to automatically segment and quantify the temporalis muscle and determine whether the temporalis muscle cross-sectional area (CSA) is able to predict survival and disease progression.

They examined baseline and follow-up cranial MR scans from patients newly diagnosed with glioblastoma between January 2015 and May 2018 at a tertiary referral centre.

The scans were taken at diagnosis, after surgery and chemoradiotherapy, and at 3-month follow-up intervals, and slices were extracted at a range of levels between the orbital roof and mid-orbit.

Following manual assessment by two trained readers to provide ‘ground truth’ labels for training and evaluation, the team trained a convolutional neural network to segment temporalis muscle images and the CSA was calculated.

One hundred and fifty-two MR scans from 45 patients were included in the analysis. The median age of the patients was 55 years, and 24.4% were female.

The team examined the performance of the neural network in segmenting the temporalis muscle using a range of loss of function assessments, with the region based Dice coefficient and the boundary based Hausdorff distance performing the best.

The average CSA error compared to human assessment was 1.94%, suggesting that the performance was comparable. However, the neural network took only 1.1 seconds to segment each image, compared to 10 minutes per case for manual segmentation.

Over an average follow-up of 19.2 months, the median overall survival was 18.3 months and the progression-free survival (PFS) was 8.9 months.

Mean temporalis CSA decreased over time, at an overall loss of 27.5 mm2 from baseline to the final follow-up (p=0.10).

Using the median baseline CSA of 507 mm2 as a cut-off, the team dichotomised the patients into low and high CSA groups.

Kaplan-Meier analysis with log rank test showed that overall survival was significantly longer with patients with high baseline CSA, at a median of 21.3 months versus 14.0 months for those with a low CSA (p=0.033).

Median PFS was also significantly longer in patients with a high CSA at baseline, at 15.1 months versus 6.1 months (p<0.0005).

Cox multivariate analysis showed that CSA was a significant independent predictor of both overall survival, at a hazard ratio of 0.380 (p=0.049), and for PFS, at a hazard ratio of 0.271 (p=0.005).

Age and, in particular, methylation of the DNA repair enzyme O[6[-methylguanine-DNA methyltransferase (MGMT) were also significant independent predictors of both overall survival and PFS, but there was no association with sex and tumour laterality.

Subgroup analysis showed that CSA was independently and significantly associated with overall survival and PFS in male patients aged under 55 years with unmethylated MGMT.

Dr Mi said the results indicate that the “temporalis muscle area is a significant prognostic marker for overall and progression-free survival in glioblastoma, corroborating previous evidence on temporalis muscle width”.

She said that the mechanism “likely reflects physical inactivity, nutritional deficiency and catabolic paraneoplastic and inflammatory processes”.

Mi believes that their deep learning approach could be used to improve “prognostic estimates and predictive models in glioblastoma” and to “optimise treatment decision-making and stratification”.

“More widely, deep learning can be used for sarcopenia screening in cancer care, and this can guide therapeutic interventions for muscle preservation, which can feasibly be deployed in clinical workflows.

“We are currently validating all of this in an external dataset,” she added.

Second Study

Presenting the second study, Dr Cheng Boon, Clatterbridge Cancer Centre, Merseyside, said that the current standard practice in prostate cancer is to use computed tomography-based prostate radiotherapy.

There is increasing interest, however, in MR imaging (MRI)-based workflow to improve prostate cancer outcomes and reduce late toxicities associated with treatment.

But an MRI-only radiotherapy workflow, Dr Boon noted, requires “daily adaptive” planning and will add “significant demands” on clinician time and expertise in oncology services.

To examine whether deep learning algorithms can be trained for radiotherapy auto-segmentation to aid prostate radiotherapy workflows, the team conducted a proof of concept study.

They used scans from three publicly available MRI prostate cancer datasets of a total of 132 patients.

The researchers applied four deep learning algorithms to the auto-segmentation of prostate GTV, training, validating and evaluating their performance against ‘ground truth’ assessments.

All four algorithms performed well, achieving an average Dice similarity coefficient of 0.89–0.92.

“We are conscious that GTV contouring is the most complicated and variable part of the clinician’s work,” Boon said, “yet with MR-based workflow this is increasingly important.”

He added that the artificial intelligence models “currently available in the market are all proprietary and are difficult to assess” as they use a “black box” approach.

In contrast, their work, which is based on open source databases, could be made “freely available and therefore can be potentially shared widely”, with the aim of allowing “future collaboration with other parties”.

The study by Mi and colleagues was funded by Imperial College London and the UK National Institute for Health Research Biomedical Centre.

The study by Gillespie and colleagues was funded by National Cancer Research Institute.

No relevant financial relationships declared.

NCRI Virtual Showcase 2020: Abstract: Deep learning-based segmentation and quantification of temporalis muscle for sarcopenia assessment is an independent prognostic factor in glioblastoma. Presented November 2.

NCRI Virtual Showcase 2020: Abstract: Deep Learning Auto-segmentation of Prostate Cancer from MRI data. Presented November 2.


Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as: