Segmentation of Pediatric Brain Tumors on Multi-Modal MRI Data using Deep Learning Approaches

Over recent years great progress has been made in the field of automated segmentation of adult brain tumors (gliomas) on multi-modal Magnetic Resonance Imaging (MRI) scans, primarily due to the success of the Multimodal Brain Tumor Image Segmentation Benchmark (BRATS) challenge [1]. However, the field of pediatric brain tumor segmentation is still a relatively unaddressed one. In contrast to adult brain tumors, pediatric brain tumors typically manifest in different regions of the brain, as well as possessing different morphological structures [2]. The segmentation of pediatric brain tumors, like their adult counterparts, is not a trivial task, and is one which requires novel additions to the best performing algorithms trained to segment adult gliomas. Furthermore, it may be necessary to retrain those models, or indeed novel ones, on a dataset of manually segmented pediatric brain tumors to achieve the desired accuracy. Alder Hey Children's Hospital is currently in possession of a wide dataset of multi-modal pediatric MRI brain scans which are currently unused, and which could form part of the training dataset for our proposed deep learning models.
The principal aim of the project will be to develop and apply existing machine learning approaches which have demonstrated great success on adult gliomas, and tailor them to address the typical characteristics of pediatric brain tumor scans on multi-modal MR images. Furthermore, it may be necessary to conceive of entirely novel network architectures to address the difference in typical morphological characteristics or histological subtypes between pediatric brain tumors and their adult counterparts. In conjunction with this, we also aim to have fully operational software, which will be built on top of the proposed machine learning pipeline, configured at Alder Hey and which will automate the segmentation of pediatric brain tumors immediately post scan, thereby saving the clinician the task of having to manually segment an MRI slice-wise.
A secondary aim will be to work closely with a clinician to tailor said software to a more appropriate format which will aim to be more clinician friendly. This will most likely involve a bounding-box approach in which a clinician can manually extract the tumorous region by drawing a rudimentary bounding box which will then be fed to the segmentation software (thereby only segmenting the region of interest and reducing computational cost). While on this note I feel compelled to add that it is of much importance to me and this project that large strides are taken in the direction of practical implementation of machine learning models/software which work in conjunction with the variability in MRI acquisition protocol across different hospitals across the world. The Brats dataset encourages challengers to compete on a very convenient volumetric multi-modal dataset. Although this challenge has yielded excellent advances in computer vison and medical imaging segmentation tasks, the practicality of applying those models to real-word scenarios, as software for hospitals, is minimal. Not all hospitals acquire brain images as 3D volumes for all modalities. And so, most of the currently existing segmentation models which succeed in reaching state of the art results on the Brats dataset would have little or no success in departments where the imaging protocol is not necessarily volumetric. A fundamental philosophical ethos of this project is to develop network architectures and proposed approaches which will account for the variability in acquisition protocols across pediatric departments.
A tertiary aim will be to work on developing novel network architectures which can compete with the current state of the art algorithms whilst paying due attention to the fundamental ethos just mentioned.

[1] Bakas, Spyridon, Mauricio Reyes, Andras Jakab, Stefan Bauer, Markus Rempfler, Alessandro Crimi, Russell Takeshi Shinohara et al. "Identifying the best machine l