Automatic location of brain tumors: expectations and reality


The precise location of the brain tumor lies in the correct interpretation of a patient’s brain scans in different modalities (MRI, PET, etc.), which allows diagnosis confirmation. While healthcare is certainly based on human competence and experience, technology seems to be advanced enough to support the decisions of health professionals with complex mathematical algorithms.

Great advances in the creation of artificial intelligence solutions with machine learning look promising to enable automation medical image analysis, thus reducing the time of image interpretation and accurately locating tumors and their subregions for the prescription of subsequent treatments and surgery planning.

However, there are many doubts about the reliability of the automated localization of brain tumors in light of the multiple challenges along the way. We decided to meet these challenges, define ways to increase the accuracy of brain tumor diagnosis automatically, and also take a look at the state of the art in this field.

The challenges of brain tumor localization

For automated brain tumor analysis, segmentation and recording are the methods that seem most difficult. Some of the following issues have not yet been resolved, and some can be addressed with a different approach to the previous steps (e.g., preprocessing).

Segmentation challenges

Segmentation allows the detection of the tumor area, including its subcomparts and surrounding tissues. Segmenting supposedly tumor brain images is a challenge on several levels:

  • If present, high-grade tumors often have unclear contours with fractures. Some tumors can also deform the surrounding tissues and present with edema or necrosis, changing the intensity of the image around the abnormality. This change can make it difficult to accurately locate the outline of the tumor, shielding and staining its edges.
  • Single-modality analysis may be insufficient to separate tumor subregions. Therefore, the combination of several modalities may be necessary to ensure a complete diagnosis. In this case, the segmentation will depend a lot on the precise preprocessing and registration stages.
  • Full contrast imbibition time and image acquisition time after contrast injection may differ, leading to significant changes in tumor appearance. It is still debatable whether there is a need to handle the unimaginable component of the injury using segmentation algorithms and, if so, how to achieve this.

Registration challenges

In the localization of brain tumors, the purpose of the registry is to allow the simultaneous analysis of different modalities in the diagnostic stage or a subsequent follow-up of tumor growth.

There are some challenges on the way to a perfect record:

  • Spacing of anisotropic voxels. Images within particular modalities are usually acquired with different anisotropic resolutions, sometimes in different orientations.
  • When the image of a patient is recorded with a healthy brain atlas, the challenge of mismatch between the two arises.

While there are a number of suggested methods for overcoming the above problems, most require significant time and computing power. This can become one of the major pitfalls on the road to solving segmentation and logging challenges.

Efficient approaches for the localization of brain tumors

Diagnosis by MRI only

To improve the accuracy of tumor location in the analysis of magnetic resonance imaging, some sequences can be recorded. The difference in the imaging results of the T1 and T2 sequences can ensure accurate automatic detection of lesions and their subregions.

For example, a T1-weighted image can be correctly segmented and therefore detect active tumor and necrosis regions, while the edema region can be segmented from a T2-weighted recorded image. When these two sequences fuse, the image analysis software algorithm is able to form the complete overview of a tumor with all the affected areas.

Cross-sectional diagnosis: magnetic resonance imaging and PET

Combined with MRI, the metabolic data of PET (blood flow, oxygen and glucose metabolism) allow you to create an accurate picture of how the tumor looks, how it is shaped and how it affects the surrounding tissue, separating the abnormality itself from the ‘edema and necrosis.

With high-grade gliomas, for example, the affected area can be deceptively extensive. But when MRI and PET images fuse, the actual division of the subregions appears.

State of the art in the automated localization of brain tumors

There are currently several approaches to managing segmentation and registration as key steps in automating brain cancer diagnosis. But they are rather seen as a set of methods where they can be partially or fully automated.

Magnetic resonance imaging medical imaging for brain tumor studies

In the survey by S. Bauer et al., the authors summarize several sets of approaches for segmenting and recording brain MRI images. Some of them can be automated, such as:

Segmentation: diffuse clustering plus knowledge-based techniques, SVM classification, difference imaging for volumetric tumor assessment, decision forests for tissue-specific segmentation, and more.

Registration: non-rigid log to capture brain displacement, geometric metamorphosis, differential analysis for tumor growth quantification, log with EM algorithm and diffusion modeling, and more.

But these are just pieces that don’t really feature an end-to-end automatic injury location system.

Automatic brain tumor segmentation using 3D convolutional neural networks

One of the latest articles (March 2016) from Stanford University presents a new one algorithm for fully automatic brain tumor segmentation based on 3D convolutional neural networks (CNN). The authors (C. Elamri and T. de Planque) claim that this algorithm achieves 89% accuracy throughout the tumor segmentation. The researchers also compared their CNN 3D method with the performance of human radiologists (85%), the leading methods of 2013 (75-82%) and 2014 (83-88%) and obtained the highest dice score.

The advantage of its approach is that it uses data analysis and machine learning to accurately detect tumor area, edema, lesions that improve and not. The authors also create tuned algorithms to process 3D images immediately without implementing 2D-oriented algorithms in a 3D environment.

This approach makes it possible to maintain accurate spatial information and improve robustness. In addition, CNN learns over time, which will further increase accuracy. However, its algorithm does not focus at all on registration, which makes it impossible to automatically locate the tumor when there are two or more modalities or sequences.


Much effort has been made to develop algorithms that allow automatic segmentation for the location of brain tumors. However, the main goal is to evolve diagnostic support software to the level of widespread clinical application, which is still a challenge. The main challenge is that radiologists and oncologists will continue to rely on manual brain tumor delimitations until there is a full-cycle option: software that performs automatic end-to-end image analysis. And most importantly, the software that doctors can use, not just researchers.

This solution should be able to know if a patient has a tumor, what tumor subregions they are, and how they are found. Later, health professionals will also need the ability to monitor tumor growth and analyze the progress of treatment (e.g., surgery, chemotherapy).

Feel free to share your thoughts in the comments.

Medical image analysis by ScienceSoft

Innovates in the prevention, diagnosis and treatment of diseases with an efficient analysis of medical images.

Source link