BrainScanCT - Technology Overview

Introduction:

Precise and early diagnosis saves lives, right now in Europe the radiology interpretations take about 2 weeks - it's often, sadly, too long. That is why we use machine learning to improve the time and efficiency of Brain CT Scans interpretations by giving radiologists an AI-based decision support system.

In the UK alone, there is a shortage of 16.000 radiologists. There is also a significant decrease in young medical professionals, choosing Radiology in relation to demand. In addition, CT scanners get cheaper and faster. The overwhelming workload produces an increased number of errors in the everyday practice of radiologists. Various studies indicate a range of 2% to 30%.

In the European Union and Switzerland, at least 52 million computed tomography (CT) examinations are performed each year [1]. Statistical surveys covering 40 countries, conducted on four continents, showed that the most often subjected to the CT examination part of the body is head [2]. The data indicate that at least 25% of CT scans are brain CTs, so it is safe to assume that about 13 million CT scans are performed every year in Europe.

At the same time, precise diagnosis and treatment of brain disease are exceptionally important. Stroke is the second most common cause of all deaths in the world (11.9%). In 2012 alone, approximately 6.7 million died due to stroke worldwide. The European Statistical Office states that in the European Union in 2013 due to cerebrovascular diseases (I60-I69) (including strokes) around 433 thousand people died, which constituted 8.7% of all deaths) [3].

BrainScan addresses those problems by developing a decision support system that reuses archived medical databases to create an AI capable of detection, classification, and localization of brain pathological changes. In a matter of seconds, the radiologist is presented with a visual representation of the likability of certain brain pathologies and a rough estimate of its localization. The data, in this specific context, could benefit the radiology workflow by speeding up the interpretation process and lowering the probability of misdiagnosis. It also suggests it’s broader clinical implementations, such as developing a feature for patient prioritization, as shown in the study published by Nature [4].

BrainScan is also developing a visual search tool, which from a healthcare-quality perspective, can give substantial value to radiologists. They can select an area of interest in the investigated CT scan, and our algorithm will search the PACS and RIS databases for.

Classification

Automatically tagging CT examinations a normal (not including pathological changes) or abnormal (including pathological changes), with the ability to diagnose 2 pathological changes now (Hemorrhage & Stroke) and up to 26 pathological changes in the near feature.

The list of the pathological changes tagged automatically by the BrainScan CT system  to the end of 2020.

  • Development and evaluation of BrainScan CT pathology detection system was conducted on a dataset 5909 brain CT scan. The dataset was split into three disjoint subsets: train set, validation set, and test set

Classification models in BrainScan CT leverage 3D volumetric MNasNet convolutional neural network architecture.

BrainScan evaluation on Qure.ai dataset

An additional evaluation was performed on CQ500 dataset of 491 scans provided by Qure.ai. The dataset contains 206 scans of a patient with hemorrhage (established by a vote of 3 radiologists). Test results are presented in Table 3.

Qure.ai evaluation on BrainScan dataset

In addition to the CQ500 dataset, we've conducted a detailed compresence where we’ve tested 16 random cases from our own dataset that were interpreted by a radiologist (8 scans containing ICH, 8 scans without ICH), which were not a part of training and development datasets in BrainScan. We evaluated both Qure.ai trial version and our own intracranial hemorrhage detection algorithm. For 14 out of 16 scans both Qure.ai and BrainScan returned correct predictions, while 2 of the scans contained subtle cases of hemorrhage and were misclassified as normal scans. Hence, evaluation on this sample yielded identical results for hemorrhage detection in both systems.

Solving the problem of discrepancy in radiology

To the best of our knowledge, BrainScan had lead the first highly comprehensive study of inter-observer agreement on 3D localization and labeling of brain lesions on computed tomography scans. Five radiologists with different levels of experience examined over 200 brain CT scans and assigned few thousand boxes denoting locations and types of brain lesions. 19 pathology types were considered in the study. Significant differences between observers' annotations were identified, even for the two most experienced radiologists (13 and 20 years of experience):

  • Highest agreement was achieved for intraventricular hemorrhage, but the result was far from perfect agreement: 92% (measured as Cohen’s Kappa coefficient);

  • Lowest agreement was achieved for cysts: 17%;

  • Ischemic strokes: 62%;

  • Tumors: 52%;

  • Average agreement for 19 pathologies: 57%;

For less experienced radiologists (2.5 years) the agreement dropped by an additional 5% on average.

In addition, the observer greatly varied in terms of the number of spotted pathologies. The most experienced radiologist assigned few times more 3D boxes than the least experienced observer.

BrainScan under-development features

Anomalies

Ability to highlight areas of the brain CT scans indicating abnormal asymmetry – as potential risk areas – and area suspected to contain pathological changes. Will enable to draw the attention of a radiologist to particular regions that outstanding normal appearance of a healthy brain.

The feature utilizes modern techniques for explaining deep learning models decisions, including:

  • Class Activation Maps
  • Guided Backpropagation)
  • Object detection algorithms
  • Siamese and triplet networks

Scan auto-alignment

Provides radiologists with a fully-automatic option to align the CT scan in both axial and coronal planes enabling clearer reading of the scan and saving precious time by eliminating the need to manually align the scan in MPR (multiplanar reconstruction) mode. Scan alignment significantly helps radiologists in analysis of symmetry of brain structures, which is a highly important part of the interpretation.

The alignment tool operates by identifying axes of symmetry of a brain, followed by applying appropriate rotations in each of the planes in order to align the axes with the view.

Search for visually similar CT scans

Enables searching databases of interpreted CT Scans (radiology atlases) for similar lesions and pathological changes including tumors, bleedings, strokes, post-stroke foci and clinically significant calcifications in soft tissues. Radiologists are therefore automatically provided with relevant reference interpretations of visually similar scans.

Search feature is based on representations learning techniques coupled with effective indexing and search algorithms.

Data acquisition for machine learning

We’ve gathered 25 thousands examinations of the brain CT scan, that gives us approximately 10M separate images (slices).

We will extend the learning set up to 60 thousands of examinations to the end of 2019 and surpass 100 thousand in 2020.

 

[1] Healthcare resource statistics - technical resources and medical technology,

[2] Vassileva J, Rehani MM, Al-Dhuhli H, Al-Naemi HM, Al-Suwaidi JS, Applegate K, et al. IAEA survey of pediatric CT practice in 40 countries in Asia, Europe, Latin America, and Africa: Part 1, frequency and appropriateness. AJR Am J Roentgenol 2012;198:1021-31

[3] Brain stroke - a growing problem in an aging society,

[4] Arbabshirani, M. R., Fornwalt, B. K., Mongelluzzo, G. J., Suever, J. D., Geise, B. D., Patel, A. A., & Moore, G. J. (2018). Advanced machine learning in action: identification of intracranial hemorrhage on computed tomography scans of the head with clinical workflow integration. npj Digital Medicine, 1(1), 9. ,